get-zeked/ai-agent-super-skill

AI Agent Builder Super-Skill

A comprehensive reference for designing, building, and deploying AI agents — from single-tool bots to production multi-agent systems — merging best practices from Claude Code's agent orchestration patterns with Perplexity Computer's deployment infrastructure.

---

Table of Contents

1. Gap Analysis Table
2. Agent Architecture & Design Patterns
3. MCP Server Development
4. RAG System Construction
5. Subagent Coordination
6. Execution Planning & Verification
7. Prompt Engineering & Optimization
8. ML Integration for Agents
9. Skill & Capability Creation
10. Backend Infrastructure for Agents
11. Agent Deployment & Monitoring
12. Unique Perplexity Computer Capabilities

---

1. Gap Analysis Table

This table maps each capability domain to its source skill, coverage level, and any gaps filled by this super-skill.

| Capability | Source Skill(s) | Coverage | Gaps Filled Here | |------------|----------------|----------|-----------------| | Agent architecture (ReAct, Plan-Execute) | senior-prompt-engineer | Partial — workflow diagrams only | Full pattern library with code | | Multi-agent orchestration | subagent-driven-development, dispatching-parallel-agents | Strong process, no code | Integration patterns, conflict detection | | MCP server building | mcp-builder | Full (4-phase process) | Perplexity-compatible CGI deployment | | RAG pipeline construction | senior-ml-engineer, senior-prompt-engineer | Chunking + DB selection tables | End-to-end pipeline code | | Prompt engineering | senior-prompt-engineer | Pattern reference table | Advanced chain-of-thought + meta-prompting | | Subagent task dispatch | subagent-driven-development | Process diagrams | Template prompts with full context injection | | Parallel agent dispatch | dispatching-parallel-agents | Decision tree + examples | Conflict detection, state isolation | | Plan execution with checkpoints | executing-plans | Step-by-step process | Batch sizing, rollback strategies | | MLOps / model deployment | senior-ml-engineer | Docker + k8s templates | Agent-specific serving patterns | | Backend webhooks/SQLite | webserver | Full CGI-bin reference | Agent memory persistence layer | | Skill packaging (SKILL.md) | create-skill (Perplexity) | YAML frontmatter format | Validation pipeline, versioning | | Deployment & observability | website-building (Perplexity) | UI deployment only | Agent health checks, trace logging | | Perplexity 400+ integrations | Perplexity Computer native | Available but undocumented | Integration mapping for agent use | | Scheduled monitoring | Perplexity Computer native | Not in any skill | Agent heartbeat and drift triggers |

---

2. Agent Architecture & Design Patterns

2.1 Core Agent Loop

Every agent follows a fundamental observe → think → act → observe loop. The differences between architectures lie in how deeply they plan before acting and how they handle tool results.

+---------------------------------------------+
|                 AGENT LOOP                  |
|                                             |
|   Input/Observation                         |
|         |                                   |
|         v                                   |
|   +-------------+                           |
|   |    Think    | <---------------------+   |
|   |  (Reason)   |                       |   |
|   +------+------+                       |   |
|          |                              |   |
|          v                              |   |
|   +-------------+    No more tools      |   |
|   |  Select     |------------------>    |   |
|   |  Action     |                  |    |   |
|   +------+------+                  |    |   |
|          |                         |    |   |
|          v                         v    |   |
|   +-------------+           +----------+|   |
|   |  Execute    |           |  Final   ||   |
|   |  Tool/API   |           |  Answer  ||   |
|   +------+------+           +----------+|   |
|          |                              |   |
|          v                              |   |
|   +-------------+                       |   |
|   |  Observe    |-----------------------+   |
|   |  Result     |                           |
|   +-------------+                           |
+---------------------------------------------+

2.2 Architecture Patterns

Pattern A: ReAct (Reason + Act)

Best for: open-ended research, customer support, tool-use agents.

How it works: The agent interleaves reasoning traces (Thought:) with actions (Action:) and observations (Observation:) in a single conversation thread.

REACT_SYSTEM_PROMPT = """
You are a research agent. For every task:

1. THOUGHT: Reason about what you know and what you need
2. ACTION: Choose one tool to call
3. OBSERVATION: Read the tool result
4. Repeat until you have enough information
5. FINAL ANSWER: Synthesize and respond

Available tools: {tool_list}

Format strictly:
Thought: <your reasoning>
Action: <tool_name>
Action Input: <tool_arguments as JSON>
Observation: <tool result — filled by system>
... (repeat)
Final Answer: <your complete response>
"""

def react_agent(query: str, tools: dict, llm, max_iterations: int = 10) -> str:
    messages = [
        {"role": "system", "content": REACT_SYSTEM_PROMPT.format(
            tool_list="\n".join(f"- {k}: {v['description']}" for k, v in tools.items())
        )},
        {"role": "user", "content": query}
    ]
    
    for iteration in range(max_iterations):
        response = llm.complete(messages)
        
        if "Final Answer:" in response:
            return response.split("Final Answer:")[-1].strip()
        
        # Parse Action / Action Input
        action_line = [l for l in response.split("\n") if l.startswith("Action:")]
        input_line  = [l for l in response.split("\n") if l.startswith("Action Input:")]
        
        if not action_line:
            break
        
        tool_name = action_line[0].replace("Action:", "").strip()
        tool_input = json.loads(input_line[0].replace("Action Input:", "").strip())
        
        # Execute tool
        if tool_name in tools:
            observation = tools[tool_name]["fn"](**tool_input)
        else:
            observation = f"Error: Unknown tool '{tool_name}'"
        
        messages.append({"role": "assistant", "content": response})
        messages.append({"role": "user", "content": f"Observation: {observation}"})
    
    return "Agent reached max iterations without a final answer."

Pattern B: Plan-and-Execute

Best for: complex multi-step workflows, code generation, structured report creation.

How it works: A planner LLM generates a complete task list first; executor agents complete each step sequentially or in parallel.

PLANNER_PROMPT = """
Given this goal: {goal}

Create a numbered execution plan. Each step must be:
- Atomic: one clear action
- Verifiable: has a concrete success criterion
- Independent (where possible): can run without other steps completing first

Output format:
PLAN:
1. [Step description] | SUCCESS: [verification criterion] | DEPS: [step numbers or NONE]
2. ...
"""

EXECUTOR_PROMPT = """
Execute this step exactly:
{step}

Context from previous steps:
{context}

Available tools: {tools}

Return:
- RESULT: what you produced
- STATUS: SUCCESS or FAILED
- NOTES: any issues or observations
"""

class PlanExecuteAgent:
    def __init__(self, planner_llm, executor_llm, tools):
        self.planner = planner_llm
        self.executor = executor_llm
        self.tools = tools
    
    def run(self, goal: str) -> dict:
        # Phase 1: Plan
        plan_response = self.planner.complete(
            PLANNER_PROMPT.format(goal=goal)
        )
        steps = self._parse_plan(plan_response)
        
        # Phase 2: Execute
        results = {}
        for step in self._topological_sort(steps):
            context = {k: v["result"] for k, v in results.items() if v["status"] == "SUCCESS"}
            result = self.executor.complete(
                EXECUTOR_PROMPT.format(
                    step=step["description"],
                    context=json.dumps(context, indent=2),
                    tools=list(self.tools.keys())
                )
            )
            results[step["id"]] = self._parse_result(result)
        
        return results

Pattern C: Reflexion

Best for: code debugging, essay writing, tasks that benefit from self-critique.

How it works: After each attempt, the agent evaluates its own output, stores a reflection in memory, and retries.

REFLEXION_EVALUATOR_PROMPT = """
Task: {task}
Attempt: {attempt}

Evaluate this attempt:
1. What did it get RIGHT? (be specific)
2. What did it get WRONG or MISS? (be specific)
3. What should the NEXT attempt do differently?

Score (0-10): 
Reflection: 
"""

class ReflexionAgent:
    def __init__(self, llm, max_attempts: int = 3, pass_threshold: float = 8.0):
        self.llm = llm
        self.max_attempts = max_attempts
        self.threshold = pass_threshold
        self.memory = []  # Persisted reflections
    
    def run(self, task: str) -> str:
        for attempt_num in range(self.max_attempts):
            # Inject prior reflections into context
            reflection_context = "\n".join(
                f"Attempt {i+1} reflection: {r}" for i, r in enumerate(self.memory)
            )
            
            attempt = self.llm.complete(
                f"Task: {task}\n\nPrior attempt learnings:\n{reflection_context}\n\nYour attempt:"
            )
            
            # Evaluate
            eval_response = self.llm.complete(
                REFLEXION_EVALUATOR_PROMPT.format(task=task, attempt=attempt)
            )
            score = float(re.search(r"Score \(0-10\):\s*([\d.]+)", eval_response).group(1))
            reflection = re.search(r"Reflection:\s*(.+)", eval_response, re.DOTALL).group(1).strip()
            
            self.memory.append(reflection)
            
            if score >= self.threshold:
                return attempt
        
        return attempt  # Return best attempt after max tries

Pattern D: Tool-Use Agent (Function Calling)

Best for: API integrations, data retrieval, modern LLM APIs that support native tool calling.

import anthropic

def build_tool_agent(tools: list[dict], system: str = "") -> callable:
    """
    tools: list of Anthropic-format tool definitions
    Returns a function that runs the agent for a given query.
    """
    client = anthropic.Anthropic()
    
    def run(query: str, tool_executors: dict[str, callable]) -> str:
        messages = [{"role": "user", "content": query}]
        
        while True:
            response = client.messages.create(
                model="claude-opus-4-5",
                max_tokens=4096,
                system=system,
                tools=tools,
                messages=messages
            )
            
            # No tool calls — final answer
            if response.stop_reason == "end_turn":
                return response.content[0].text
            
            # Process tool calls
            tool_results = []
            for block in response.content:
                if block.type == "tool_use":
                    executor = tool_executors.get(block.name)
                    if executor:
                        result = executor(**block.input)
                        tool_results.append({
                            "type": "tool_result",
                            "tool_use_id": block.id,
                            "content": str(result)
                        })
            
            messages.append({"role": "assistant", "content": response.content})
            messages.append({"role": "user", "content": tool_results})
    
    return run

2.3 Architecture Selection Guide

| Goal | Pattern | Reason | |------|---------|--------| | Open-ended research | ReAct | Flexible, self-correcting, handles unknown paths | | Multi-step report generation | Plan-Execute | Predictable, auditable, checkpointable | | Code writing / debugging | Reflexion | Self-critique loop improves quality over iterations | | API integration / tool calling | Tool-Use | Native LLM feature, lower latency, less prompt engineering | | Customer support bot | ReAct + Tool-Use | Hybrid: structured tools with flexible reasoning | | Batch data processing | Plan-Execute with parallel dispatch | Speed via parallelism, structured output | | Creative tasks (writing, design) | Reflexion | Quality improves with each self-critique cycle |

2.4 Multi-Agent System Topologies

TOPOLOGY 1: Hub-and-Spoke (Orchestrator + Specialists)

                    +----------------+
                    |  Orchestrator  |
                    |  (Coordinator) |
                    +-------+--------+
          +-----------------+-----------------+
          v                 v                 v
   +------------+   +------------+   +------------+
   |  Research  |   |   Coder    |   |  Writer    |
   |   Agent    |   |   Agent    |   |   Agent    |
   +------------+   +------------+   +------------+

Use for: complex tasks needing specialized expertise per sub-domain.
Orchestrator decomposes goal -> dispatches -> aggregates results.

TOPOLOGY 2: Pipeline (Assembly Line)

  Input -> [Extractor] -> [Transformer] -> [Validator] -> [Writer] -> Output

Use for: ETL, document processing, multi-stage generation tasks.
Each agent only sees the previous stage's output.

TOPOLOGY 3: Competitive / Debate

  Query -> Agent A --+
                    +---> Judge Agent ---> Final Answer
  Query -> Agent B --+

Use for: decisions requiring multiple perspectives, factual verification,
high-stakes outputs where consensus improves reliability.

TOPOLOGY 4: Peer Network (Gossip/Consensus)

  Agent 1 <---> Agent 2
     ^  \          ^
     |   \         |
     v    \        v
  Agent 4 <---> Agent 3

Use for: simulation, emergent behavior research, distributed problem solving.
High coordination overhead — avoid for production automation.

---

3. MCP Server Development

3.1 Four-Phase MCP Build Process

Building a production MCP server follows four phases. Do not skip phases — each builds on the previous.

Phase 1: Research & Planning

1. Fetch MCP spec: https://modelcontextprotocol.io/sitemap.xml then pages with .md suffix
2. Study the target API's documentation — auth requirements, rate limits, key endpoints
3. Decide: TypeScript (recommended) or Python (FastMCP)
4. List all tools, prioritizing comprehensive API coverage over convenience wrappers

Phase 2: Implementation

Phase 3: Review & Test — use MCP Inspector

Phase 4: Create Evaluations — 10 read-only, complex, verifiable questions

3.2 TypeScript MCP Server Template

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

// --- Server Initialization ---------------------------------------------------
const server = new McpServer({
  name: "my-service-mcp",
  version: "1.0.0",
});

// --- Shared API Client --------------------------------------------------------
interface ApiConfig {
  baseUrl: string;
  apiKey: string;
}

class ServiceClient {
  constructor(private config: ApiConfig) {}

  async request<T>(
    endpoint: string,
    options: RequestInit = {}
  ): Promise<T> {
    const url = `${this.config.baseUrl}${endpoint}`;
    const response = await fetch(url, {
      ...options,
      headers: {
        "Authorization": `Bearer ${this.config.apiKey}`,
        "Content-Type": "application/json",
        ...options.headers,
      },
    });

    if (!response.ok) {
      const error = await response.text();
      throw new Error(
        `API error ${response.status}: ${error}. ` +
        `Suggestion: Check your API key or verify the resource exists.`
      );
    }

    return response.json() as Promise<T>;
  }
}

const client = new ServiceClient({
  baseUrl: process.env.SERVICE_BASE_URL ?? "https://api.example.com",
  apiKey: process.env.SERVICE_API_KEY ?? "",
});

// --- Tool: List Items ---------------------------------------------------------
server.registerTool(
  "service_list_items",
  {
    description: "List items with optional filtering and pagination.",
    inputSchema: z.object({
      page:     z.number().int().min(1).default(1).describe("Page number (1-indexed)"),
      per_page: z.number().int().min(1).max(100).default(20).describe("Items per page"),
      filter:   z.string().optional().describe("Optional keyword filter"),
    }),
    annotations: {
      readOnlyHint:  true,
      destructiveHint: false,
      idempotentHint: true,
    },
  },
  async (params) => {
    const query = new URLSearchParams({
      page:     String(params.page),
      per_page: String(params.per_page),
      ...(params.filter ? { q: params.filter } : {}),
    });

    const data = await client.request<{ items: unknown[]; total: number }>(
      `/items?${query}`
    );

    return {
      content: [{
        type: "text",
        text: JSON.stringify(data, null, 2),
      }],
      structuredContent: data,
    };
  }
);

// --- Tool: Get Item -----------------------------------------------------------
server.registerTool(
  "service_get_item",
  {
    description: "Get a single item by ID.",
    inputSchema: z.object({
      id: z.string().describe("Item ID"),
    }),
    annotations: {
      readOnlyHint: true,
      destructiveHint: false,
      idempotentHint: true,
    },
  },
  async (params) => {
    const item = await client.request(`/items/${params.id}`);
    return {
      content: [{ type: "text", text: JSON.stringify(item, null, 2) }],
      structuredContent: item,
    };
  }
);

// --- Tool: Create Item --------------------------------------------------------
server.registerTool(
  "service_create_item",
  {
    description: "Create a new item. Returns the created item with its assigned ID.",
    inputSchema: z.object({
      name:        z.string().min(1).describe("Item name"),
      description: z.string().optional().describe("Optional item description"),
      tags:        z.array(z.string()).optional().describe("Tag list"),
    }),
    annotations: {
      readOnlyHint:    false,
      destructiveHint: false,
      idempotentHint:  false,
    },
  },
  async (params) => {
    const item = await client.request("/items", {
      method: "POST",
      body: JSON.stringify(params),
    });
    return {
      content: [{ type: "text", text: JSON.stringify(item, null, 2) }],
      structuredContent: item,
    };
  }
);

// --- Transport ----------------------------------------------------------------
const transport = new StdioServerTransport();
await server.connect(transport);
console.error("MCP server running on stdio");

3.3 Python FastMCP Server Template

#!/usr/bin/env python3
"""FastMCP server template for Python-based MCP servers."""

import json
import os
from typing import Any

import httpx
from fastmcp import FastMCP
from pydantic import BaseModel, Field

# --- Server Initialization ---------------------------------------------------
mcp = FastMCP("my-service-mcp")

# --- API Client ---------------------------------------------------------------
BASE_URL = os.environ.get("SERVICE_BASE_URL", "https://api.example.com")
API_KEY  = os.environ.get("SERVICE_API_KEY", "")

async def api_request(endpoint: str, method: str = "GET", body: dict | None = None) -> Any:
    headers = {
        "Authorization": f"Bearer {API_KEY}",
        "Content-Type": "application/json",
    }
    async with httpx.AsyncClient() as client:
        response = await client.request(
            method,
            f"{BASE_URL}{endpoint}",
            headers=headers,
            json=body,
        )
        if not response.is_success:
            raise ValueError(
                f"API error {response.status_code}: {response.text}. "
                "Check your credentials or verify the resource exists."
            )
        return response.json()

# --- Input Models -------------------------------------------------------------
class ListParams(BaseModel):
    page:     int   = Field(default=1,  ge=1,          description="Page number (1-indexed)")
    per_page: int   = Field(default=20, ge=1,  le=100, description="Items per page")
    filter:   str | None = Field(default=None,         description="Keyword filter")

class CreateParams(BaseModel):
    name:        str        = Field(...,       description="Item name")
    description: str | None = Field(None,      description="Optional description")
    tags:        list[str]  = Field(default=[], description="Tag list")

# --- Tools --------------------------------------------------------------------
@mcp.tool(description="List items with optional filtering and pagination.")
async def service_list_items(params: ListParams) -> str:
    query = f"?page={params.page}&per_page={params.per_page}"
    if params.filter:
        query += f"&q={params.filter}"
    data = await api_request(f"/items{query}")
    return json.dumps(data, indent=2)


@mcp.tool(description="Get a single item by ID.")
async def service_get_item(id: str) -> str:
    """id: Item ID to retrieve"""
    item = await api_request(f"/items/{id}")
    return json.dumps(item, indent=2)


@mcp.tool(description="Create a new item. Returns the created item with its assigned ID.")
async def service_create_item(params: CreateParams) -> str:
    item = await api_request("/items", method="POST", body=params.model_dump())
    return json.dumps(item, indent=2)


if __name__ == "__main__":
    mcp.run()

3.4 MCP Tool Design Checklist

Before shipping any MCP server, verify every tool against this checklist:

• [ ] Tool name uses service_verb_noun convention (e.g., github_create_issue)
• [ ] Description is a single sentence — concise, action-oriented
• [ ] All input fields have description populated
• [ ] Required vs. optional fields are correctly marked
• [ ] Numeric fields have min/max constraints
• [ ] Enum fields use z.enum()/Literal instead of free strings
• [ ] Annotations set: readOnlyHint, destructiveHint, idempotentHint
• [ ] Error messages suggest a remediation action
• [ ] Pagination supported for list endpoints
• [ ] structuredContent returned alongside text content
• [ ] Build compiles without errors: npm run build or python -m py_compile
• [ ] Tested with MCP Inspector: npx @modelcontextprotocol/inspector

3.5 MCP Error Message Patterns

Good error messages are diagnostic and actionable:

// Bad
throw new Error("Not found");

// Good
throw new Error(
  `Item '${id}' not found. ` +
  `Use service_list_items to find valid IDs, or verify the item exists in the service.`
);

// Bad
throw new Error("Unauthorized");

// Good
throw new Error(
  `Authentication failed. ` +
  `Check that SERVICE_API_KEY is set and has the required 'items:read' scope. ` +
  `Generate a new key at https://service.example.com/settings/api-keys`
);

---

4. RAG System Construction

4.1 Complete RAG Pipeline

+---------------------------------------------------------------------+
|                       INGESTION PIPELINE                            |
|                                                                     |
|  Documents -> [Loader] -> [Chunker] -> [Embedder] -> [Vector Store] |
|                                          |                          |
|                                   [Metadata Store]                  |
+---------------------------------------------------------------------+
                                    |
                              (index built)
                                    |
+---------------------------------------------------------------------+
|                        QUERY PIPELINE                               |
|                                                                     |
|  Query -> [Query Embed] -> [Vector Search] -> [Reranker] -> [LLM]  |
|              |                  |                |                  |
|         [HyDE opt.]        [Metadata          [Context             |
|                              Filter]           Format]              |
+---------------------------------------------------------------------+

4.2 Full Python RAG Implementation

#!/usr/bin/env python3
"""
Production RAG pipeline with chunking, embedding, retrieval, and reranking.
"""

import hashlib
import json
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any

# --- Data Models -------------------------------------------------------------
@dataclass
class Document:
    content: str
    metadata: dict[str, Any] = field(default_factory=dict)
    doc_id: str = field(default="")
    
    def __post_init__(self):
        if not self.doc_id:
            self.doc_id = hashlib.md5(self.content.encode()).hexdigest()[:12]

@dataclass
class Chunk:
    content: str
    doc_id: str
    chunk_index: int
    metadata: dict[str, Any] = field(default_factory=dict)
    embedding: list[float] = field(default_factory=list)

@dataclass
class RetrievedChunk:
    chunk: Chunk
    score: float
    rerank_score: float = 0.0

# --- Chunking Strategies -----------------------------------------------------
class ChunkingStrategy:
    """Base class — override chunk()."""
    def chunk(self, doc: Document) -> list[Chunk]:
        raise NotImplementedError

class FixedSizeChunker(ChunkingStrategy):
    """Fixed token-count chunks with overlap. Good for general text."""
    
    def __init__(self, chunk_size: int = 512, overlap: int = 64):
        self.chunk_size = chunk_size
        self.overlap    = overlap
    
    def chunk(self, doc: Document) -> list[Chunk]:
        # Simple word-based split (use tiktoken for production token counting)
        words = doc.content.split()
        chunks = []
        step = self.chunk_size - self.overlap
        
        for i in range(0, len(words), step):
            chunk_words = words[i:i + self.chunk_size]
            if not chunk_words:
                break
            chunks.append(Chunk(
                content=" ".join(chunk_words),
                doc_id=doc.doc_id,
                chunk_index=len(chunks),
                metadata=doc.metadata,
            ))
        return chunks

class SentenceChunker(ChunkingStrategy):
    """Sentence-boundary-aware chunking. Better for structured prose."""
    
    def __init__(self, sentences_per_chunk: int = 5, overlap_sentences: int = 1):
        self.n  = sentences_per_chunk
        self.ov = overlap_sentences
    
    def chunk(self, doc: Document) -> list[Chunk]:
        import re
        sentences = re.split(r'(?<=[.!?])\s+', doc.content.strip())
        chunks = []
        step = self.n - self.ov
        
        for i in range(0, len(sentences), step):
            batch = sentences[i:i + self.n]
            if not batch:
                break
            chunks.append(Chunk(
                content=" ".join(batch),
                doc_id=doc.doc_id,
                chunk_index=len(chunks),
                metadata=doc.metadata,
            ))
        return chunks

class RecursiveChunker(ChunkingStrategy):
    """Hierarchical chunking — preserves section structure. Best for long docs."""
    
    SEPARATORS = ["\n\n", "\n", ". ", " "]
    
    def __init__(self, max_chunk_size: int = 800, min_chunk_size: int = 100):
        self.max_size = max_chunk_size
        self.min_size = min_chunk_size
    
    def chunk(self, doc: Document) -> list[Chunk]:
        chunks = []
        self._split(doc.content, doc.doc_id, doc.metadata, 0, chunks, [])
        return chunks
    
    def _split(self, text, doc_id, metadata, depth, chunks, idx_counter):
        if len(text.split()) <= self.max_size or depth >= len(self.SEPARATORS):
            idx_counter.append(None)
            chunks.append(Chunk(
                content=text,
                doc_id=doc_id,
                chunk_index=len(chunks),
                metadata=metadata,
            ))
            return
        
        sep = self.SEPARATORS[depth]
        parts = text.split(sep)
        current = ""
        
        for part in parts:
            candidate = (current + sep + part).strip() if current else part
            if len(candidate.split()) <= self.max_size:
                current = candidate
            else:
                if current and len(current.split()) >= self.min_size:
                    self._split(current, doc_id, metadata, depth + 1, chunks, idx_counter)
                current = part
        
        if current:
            self._split(current, doc_id, metadata, depth + 1, chunks, idx_counter)

# --- Vector Store (using Chroma for local dev) --------------------------------
class VectorStore:
    """Minimal abstract vector store interface."""
    
    def upsert(self, chunks: list[Chunk]) -> None:
        raise NotImplementedError
    
    def query(self, embedding: list[float], top_k: int = 10,
              filter_metadata: dict | None = None) -> list[RetrievedChunk]:
        raise NotImplementedError

class ChromaVectorStore(VectorStore):
    """Local development store using Chroma."""
    
    def __init__(self, collection_name: str, persist_dir: str = "./chroma_db"):
        import chromadb
        client = chromadb.PersistentClient(path=persist_dir)
        self.collection = client.get_or_create_collection(
            name=collection_name,
            metadata={"hnsw:space": "cosine"},
        )
    
    def upsert(self, chunks: list[Chunk]) -> None:
        self.collection.upsert(
            ids=[f"{c.doc_id}_{c.chunk_index}" for c in chunks],
            documents=[c.content for c in chunks],
            embeddings=[c.embedding for c in chunks],
            metadatas=[c.metadata for c in chunks],
        )
    
    def query(self, embedding: list[float], top_k: int = 10,
              filter_metadata: dict | None = None) -> list[RetrievedChunk]:
        kwargs: dict[str, Any] = {
            "query_embeddings": [embedding],
            "n_results": top_k,
            "include": ["documents", "metadatas", "distances"],
        }
        if filter_metadata:
            kwargs["where"] = filter_metadata
        
        results = self.collection.query(**kwargs)
        retrieved = []
        for doc, meta, dist in zip(
            results["documents"][0],
            results["metadatas"][0],
            results["distances"][0],
        ):
            retrieved.append(RetrievedChunk(
                chunk=Chunk(content=doc, doc_id=meta.get("doc_id",""), chunk_index=0, metadata=meta),
                score=1.0 - dist,  # cosine: distance -> similarity
            ))
        return retrieved

# --- Embedder ----------------------------------------------------------------
class Embedder:
    """Embed text using OpenAI-compatible API."""
    
    def __init__(self, model: str = "text-embedding-3-small"):
        import openai
        self.client = openai.OpenAI()
        self.model  = model
    
    def embed(self, texts: list[str]) -> list[list[float]]:
        response = self.client.embeddings.create(model=self.model, input=texts)
        return [item.embedding for item in response.data]
    
    def embed_one(self, text: str) -> list[float]:
        return self.embed([text])[0]

# --- Reranker ----------------------------------------------------------------
class CrossEncoderReranker:
    """Rerank retrieved chunks with a cross-encoder for precision improvement."""
    
    def __init__(self, model: str = "cross-encoder/ms-marco-MiniLM-L-6-v2"):
        from sentence_transformers import CrossEncoder
        self.model = CrossEncoder(model)
    
    def rerank(self, query: str, chunks: list[RetrievedChunk], top_k: int = 5) -> list[RetrievedChunk]:
        pairs  = [(query, c.chunk.content) for c in chunks]
        scores = self.model.predict(pairs)
        for chunk, score in zip(chunks, scores):
            chunk.rerank_score = float(score)
        return sorted(chunks, key=lambda c: c.rerank_score, reverse=True)[:top_k]

# --- RAG Pipeline ------------------------------------------------------------
class RAGPipeline:
    def __init__(
        self,
        chunker: ChunkingStrategy,
        embedder: Embedder,
        vector_store: VectorStore,
        reranker: CrossEncoderReranker | None = None,
        llm_fn: callable = None,
        retrieval_top_k: int = 10,
        rerank_top_k: int = 5,
    ):
        self.chunker      = chunker
        self.embedder     = embedder
        self.store        = vector_store
        self.reranker     = reranker
        self.llm          = llm_fn
        self.retrieval_k  = retrieval_top_k
        self.rerank_k     = rerank_top_k
    
    # -- Ingestion ------------------------------------------------------------
    def ingest(self, documents: list[Document]) -> int:
        all_chunks: list[Chunk] = []
        
        for doc in documents:
            chunks = self.chunker.chunk(doc)
            texts  = [c.content for c in chunks]
            embeddings = self.embedder.embed(texts)
            for chunk, emb in zip(chunks, embeddings):
                chunk.embedding = emb
            all_chunks.extend(chunks)
        
        self.store.upsert(all_chunks)
        return len(all_chunks)
    
    # -- Query -----------------------------------------------------------------
    def retrieve(
        self,
        query: str,
        metadata_filter: dict | None = None,
        use_hyde: bool = False,
    ) -> list[RetrievedChunk]:
        # Optional HyDE: generate a hypothetical document to improve retrieval
        embed_target = query
        if use_hyde and self.llm:
            hypothetical = self.llm(
                f"Write a short passage that would answer this question:\n{query}"
            )
            embed_target = hypothetical
        
        q_embedding = self.embedder.embed_one(embed_target)
        candidates  = self.store.query(q_embedding, top_k=self.retrieval_k,
                                       filter_metadata=metadata_filter)
        
        if self.reranker and candidates:
            return self.reranker.rerank(query, candidates, top_k=self.rerank_k)
        return candidates[:self.rerank_k]
    
    def answer(self, query: str, metadata_filter: dict | None = None) -> dict:
        chunks = self.retrieve(query, metadata_filter=metadata_filter)
        
        context = "\n\n---\n\n".join(
            f"[Source {i+1}] {c.chunk.content}"
            for i, c in enumerate(chunks)
        )
        
        prompt = f"""Answer the question using ONLY the provided context.
If the context doesn't contain the answer, say "I don't have enough information."

Context:
{context}

Question: {query}

Answer:"""
        
        response = self.llm(prompt) if self.llm else "[No LLM configured]"
        
        return {
            "answer": response,
            "sources": [
                {
                    "content":  c.chunk.content[:200] + "...",
                    "score":    c.score,
                    "rerank":   c.rerank_score,
                    "metadata": c.chunk.metadata,
                }
                for c in chunks
            ],
        }

4.3 Vector Database Selection

| Database | Hosting | Scale | Latency | Best For | |----------|---------|-------|---------|----------| | Pinecone | Managed cloud | High | Low | Production, zero-ops | | Qdrant | Self-hosted / cloud | High | Very Low | Performance-critical | | Weaviate | Both | High | Low | Hybrid (keyword + vector) | | Chroma | Self-hosted | Medium | Low | Local dev / prototyping | | pgvector | Self-hosted (Postgres) | Medium | Medium | Existing Postgres stacks | | Redis VSS | Both | Medium | Very Low | Real-time / cache-adjacent | | Milvus | Self-hosted / cloud | Very High | Low | Enterprise scale |

4.4 Chunking Strategy Selection

| Strategy | Chunk Size | Overlap | Best For | |----------|------------|---------|----------| | Fixed-size | 500-1000 tokens | 50-100 tokens | General text, unknown structure | | Sentence | 3-5 sentences | 1 sentence | News articles, documentation | | Semantic | Variable | Meaning-based | Research papers, books | | Recursive | Hierarchical | Parent-child | Long documents with headers |

4.5 RAG Evaluation Metrics

| Metric | Definition | Target | |--------|------------|--------| | Context Relevance | % of retrieved chunks relevant to query | > 0.80 | | Answer Faithfulness | % of answer grounded in context | > 0.90 | | Retrieval Precision@5 | Relevant chunks in top 5 / 5 | > 0.70 | | Context Coverage | % of questions with >=1 relevant chunk in top-5 | > 0.85 | | End-to-end Accuracy | Correct answers / total questions | > 0.80 |

# Evaluate a RAG pipeline
python scripts/rag_evaluator.py \
  --contexts retrieved_contexts.json \
  --questions eval_questions.json \
  --metrics relevance,faithfulness,coverage \
  --output report.json

---

5. Subagent Coordination

5.1 Subagent-Driven Development

Core principle: Fresh subagent per task + two-stage review (spec compliance, then code quality) = high quality, fast iteration.

This pattern runs entirely within the current session — no context switch to parallel sessions.

PROCESS FLOW:

1. Read plan -> extract all tasks with full text -> create TodoList

2. FOR EACH TASK:
   a. Dispatch Implementer subagent (full task text + context injected)
      +-> Subagent asks questions? -> Answer -> Re-dispatch
      +-> Subagent implements, tests, self-reviews, signals done
   
   b. Dispatch Spec Compliance Reviewer
      +-> Reviewer finds issues? -> Implementer fixes -> Re-review
      +-> OK Spec compliant -> proceed
   
   c. Dispatch Code Quality Reviewer (ONLY after spec review passes)
      +-> Reviewer finds issues? -> Implementer fixes -> Re-review
      +-> OK Quality approved -> mark task complete
   
3. After all tasks: Dispatch Final Code Reviewer for full implementation

4. Use finishing-a-development-branch workflow

5.2 Implementer Subagent Prompt Template

# Implementer Subagent

## Context
You are implementing one task from a larger plan. You have been given full task text below.
Do NOT read plan files — the controller has already provided all necessary context.

## Project Context
{project_description}
Repository: {repo_path}
Branch: {branch_name}
Tech stack: {stack}

## Your Task
{full_task_text}

## Requirements
1. Ask questions BEFORE beginning if anything is unclear
2. Follow TDD: write failing test first, then implementation
3. Run all tests and verify they pass
4. Self-review: check for edge cases, naming, error handling
5. Commit with a descriptive message

## Output When Done
- Summary of what you implemented
- Test results (pass/fail counts)
- Any concerns or trade-offs you made
- Commit SHA

5.3 Spec Reviewer Prompt Template

# Spec Compliance Reviewer

## Your Role
You are a spec compliance reviewer — NOT a code quality reviewer.
Your ONLY job: verify the implementation matches the spec. Nothing more.

## Task Spec
{task_spec}

## Implementation to Review
Git SHAs of new commits: {commit_shas}

## Review Criteria
Check for:
1. MISSING: Requirements in the spec not implemented
2. EXTRA: Features implemented that were NOT requested (scope creep)
3. WRONG: Implementation that contradicts the spec

## Output Format
STATUS: OK COMPLIANT or NOT COMPLIANT

If non-compliant, list each issue as:
- MISSING: [description]
- EXTRA: [description]  
- WRONG: [description]

Do NOT comment on code quality, style, or performance.

5.4 Code Quality Reviewer Prompt Template

# Code Quality Reviewer

## Your Role
You are a code quality reviewer. The spec compliance reviewer has already confirmed
this implementation matches the spec — your job is code quality ONLY.

## Implementation to Review
Git SHAs: {commit_shas}

## Review Criteria
For each finding, classify as:
- CRITICAL: Must fix before merge (security, correctness, data loss)
- IMPORTANT: Should fix (maintainability, performance)
- SUGGESTION: Nice to have (style, naming)

## What to Check
- Error handling completeness
- Edge cases (null, empty, boundary values)
- Naming clarity
- Magic numbers/strings (extract to constants)
- DRY violations
- Security issues (injection, auth bypass, data exposure)
- Test coverage adequacy

## Output Format
STRENGTHS: [what's well done]
CRITICAL ISSUES: [list or "None"]
IMPORTANT ISSUES: [list or "None"]
VERDICT: OK APPROVED or CHANGES REQUIRED

5.5 Red Flags in Subagent Coordination

Never do these:

• Start code quality review before spec compliance passes — wrong order produces wasted cycles
• Dispatch multiple implementer subagents in parallel on the same codebase — merge conflicts guaranteed
• Let subagent read plan files — provide full task text in the prompt instead (eliminates file-reading overhead)
• Accept "close enough" spec compliance — reviewer found issues means the task is not done
• Skip the re-review after fixes — don't trust the fix without verification
• Skip scene-setting context in subagent prompts — subagent needs to understand where the task fits

---

6. Execution Planning & Verification

6.1 Parallel Agent Dispatch Pattern

Use when 2+ independent tasks can proceed without shared state or sequential dependencies.

Decision tree:

Multiple independent tasks?
+-- YES: Can they write to the same files/resources?
|   +-- YES -> Sequential agents (avoid conflict)
|   +-- NO  -> Parallel dispatch OK
+-- NO: Tasks are related -> Single agent investigates all

Parallel dispatch template:

import asyncio
from typing import Callable, Any

async def dispatch_parallel_agents(
    tasks: list[dict],
    agent_fn: Callable[[dict], Any],
    max_concurrent: int = 5,
) -> list[dict]:
    """
    Dispatch multiple agent tasks in parallel with a concurrency limit.
    
    tasks: list of dicts, each with 'id', 'description', 'context', 'constraints'
    agent_fn: async function(task) -> result
    """
    semaphore = asyncio.Semaphore(max_concurrent)
    
    async def run_with_semaphore(task: dict) -> dict:
        async with semaphore:
            try:
                result = await agent_fn(task)
                return {"task_id": task["id"], "status": "success", "result": result}
            except Exception as e:
                return {"task_id": task["id"], "status": "failed", "error": str(e)}
    
    results = await asyncio.gather(*[run_with_semaphore(t) for t in tasks])
    return list(results)


def check_result_conflicts(results: list[dict]) -> list[str]:
    """
    Scan parallel agent results for potential conflicts before integration.
    Checks: same files modified, same database records mutated.
    """
    conflicts = []
    files_modified = {}
    
    for result in results:
        if result["status"] != "success":
            continue
        files = result.get("result", {}).get("files_modified", [])
        for f in files:
            if f in files_modified:
                conflicts.append(
                    f"Conflict: '{f}' modified by both task "
                    f"{files_modified[f]} and {result['task_id']}"
                )
            else:
                files_modified[f] = result["task_id"]
    
    return conflicts

6.2 Focused Agent Task Prompt Structure

Good parallel agent task prompts are self-contained, specific, and constrained:

## Task: {task_title}

### Problem Statement
{specific_error_messages_or_failure_description}

### Scope
Files/subsystems in scope: {explicit_list}
Files/subsystems OUT of scope: {explicit_exclusions}

### Goal
{single_clear_success_criterion}

### Constraints
- Do NOT modify: {protected_files}
- Do NOT add new dependencies without flagging it
- {other_constraints}

### Required Output
Return:
1. Root cause analysis
2. What you changed and why
3. Verification output (test results, logs)
4. Summary of changes as a git diff or commit SHA

6.3 Batch Execution with Checkpoints

EXECUTING-PLANS PROCESS:

1. LOAD & REVIEW
   - Read plan file once
   - Identify concerns or blockers -> raise with human BEFORE starting
   - Create TodoList from all tasks
   - Announce: "Using executing-plans to implement this plan."

2. EXECUTE BATCH (default: 3 tasks per batch)
   For each task in batch:
   - Mark in_progress
   - Follow steps exactly as written
   - Run all verifications specified in plan
   - Mark completed

3. CHECKPOINT REPORT
   - Show: what was implemented, verification output
   - Say: "Ready for feedback."
   - Wait for approval before next batch

4. REPEAT until all tasks complete

5. FINISH
   - Use finishing-a-development-branch workflow
   - Verify all tests pass, no regressions

STOP IMMEDIATELY if:
- A blocker appears mid-batch
- Verification fails repeatedly
- Instructions are ambiguous
-> Ask for clarification; never guess.

6.4 Verification Workflow

Every implementation task should define explicit verification steps:

VERIFICATION_LEVELS = {
    "smoke": [
        "Application starts without errors",
        "All previously-passing tests still pass",
        "No new lint errors",
    ],
    "functional": [
        "New unit tests written and passing",
        "Integration tests pass against staging",
        "Edge cases tested (null, empty, boundary)",
    ],
    "acceptance": [
        "Feature works end-to-end as specified",
        "Performance within targets (latency, throughput)",
        "Security review passed",
        "Documentation updated",
    ],
}

def build_verification_prompt(task: str, level: str = "functional") -> str:
    checks = "\n".join(f"- [ ] {c}" for c in VERIFICATION_LEVELS[level])
    return f"""
After implementing: {task}

Run these verification checks:
{checks}

Report each check as OK PASS, FAIL, or SKIPPED (with reason).
If any check FAILS, stop and report before proceeding.
"""

---

7. Prompt Engineering & Optimization

7.1 Core Prompt Patterns

| Pattern | When to Use | Token Cost | Quality Gain | |---------|-------------|------------|-------------| | Zero-shot | Simple, well-defined tasks | Lowest | Baseline | | Few-shot (3-5 examples) | Complex tasks, consistent format needed | Medium | High | | Chain-of-Thought (CoT) | Reasoning, math, multi-step logic | Medium | High | | Role Prompting | Domain expertise, specific perspective | Low | Medium | | Structured Output | Need parseable JSON/XML | Low | High (reliability) | | Tree-of-Thought | Complex problem solving, backtracking | High | Very High | | Meta-prompting | Generating/optimizing other prompts | High | Very High | | Self-consistency | High-stakes decisions (majority vote) | Very High | High |

7.2 Chain-of-Thought Implementation

COT_TEMPLATES = {
    # Standard CoT
    "standard": """
{task}

Think step by step:
1. First, identify what information is given
2. Determine what needs to be found
3. Work through the reasoning systematically
4. State your conclusion

Reasoning:
""",

    # Few-shot CoT
    "few_shot": """
Solve problems by thinking step by step.

Example 1:
Problem: {example_problem_1}
Reasoning: {example_reasoning_1}
Answer: {example_answer_1}

Example 2:
Problem: {example_problem_2}
Reasoning: {example_reasoning_2}
Answer: {example_answer_2}

Now solve:
Problem: {problem}
Reasoning:
""",

    # Zero-shot CoT (Kojima et al.)
    "zero_shot": "{task}\n\nLet's think step by step.",

    # Plan-then-execute CoT
    "plan_execute": """
{task}

Step 1 - Make a plan: List the sub-problems you need to solve, in order.
Step 2 - Execute: Work through each sub-problem, showing your reasoning.
Step 3 - Verify: Check your answer against the original question.

Begin:
""",
}

7.3 Structured Output Design

from typing import Literal
from pydantic import BaseModel, Field

# --- Define Output Schema -----------------------------------------------------
class SentimentAnalysis(BaseModel):
    summary:    str     = Field(..., max_length=200, description="Brief content summary")
    sentiment:  Literal["positive", "negative", "neutral", "mixed"]
    confidence: float   = Field(..., ge=0.0, le=1.0, description="Confidence 0-1")
    key_points: list[str] = Field(..., max_items=5, description="Up to 5 key points")

# --- Build Prompt with Schema -------------------------------------------------
def build_structured_prompt(content: str, schema: type[BaseModel]) -> str:
    schema_json = schema.model_json_schema()
    
    return f"""Analyze the following content.

Respond ONLY with valid JSON matching this schema:
{json.dumps(schema_json, indent=2)}

IMPORTANT:
- Start your response with {{
- End your response with }}
- No markdown code fences, no explanation outside the JSON

Content to analyze:
{content}

JSON response:"""

# --- Parse and Validate Output ------------------------------------------------
def parse_structured_output(response: str, schema: type[BaseModel]) -> BaseModel:
    # Strip markdown fences if present
    import re
    cleaned = re.sub(r"```(?:json)?\s*|\s*```", "", response).strip()
    
    # Find outermost JSON object
    start = cleaned.find("{")
    end   = cleaned.rfind("}") + 1
    if start == -1 or end == 0:
        raise ValueError(f"No JSON object found in response:\n{response}")
    
    data = json.loads(cleaned[start:end])
    return schema.model_validate(data)

7.4 Prompt Optimization Workflow

STEP 1: Baseline
  python scripts/prompt_optimizer.py current_prompt.txt --analyze --output baseline.json
  Capture: token count, clarity score, issues found

STEP 2: Identify Problems
  | Issue              | Apply This Pattern              |
  |--------------------|----------------------------------|
  | Ambiguous output   | Add explicit format specification |
  | Too verbose        | Extract to few-shot examples      |
  | Inconsistent results| Add role/persona framing         |
  | Missing edge cases | Add constraint boundaries         |
  | Poor reasoning     | Add chain-of-thought trigger      |
  | Wrong format       | Add schema + format enforcement   |

STEP 3: Apply Optimizations
  python scripts/prompt_optimizer.py current_prompt.txt --optimize --output optimized.txt

STEP 4: Compare
  python scripts/prompt_optimizer.py optimized.txt --analyze --compare baseline.json

STEP 5: A/B Test
  Run both prompts against held-out evaluation set.
  Accept optimization only if: quality up AND cost <= 1.2x baseline.

7.5 Meta-Prompting (Prompt Generation)

META_PROMPT_GENERATOR = """
You are an expert prompt engineer. Generate an optimized prompt for the following use case.

## Use Case
Task: {task_description}
Model: {model}
Expected input format: {input_format}
Expected output format: {output_format}
Edge cases to handle: {edge_cases}
Constraints: {constraints}

## Generate a prompt that:
1. Uses role framing appropriate for the task
2. Provides clear, unambiguous instructions
3. Includes 2-3 few-shot examples if appropriate
4. Specifies exact output format
5. Handles the listed edge cases
6. Is token-efficient (no redundancy)

Return the complete prompt, ready to use.
"""

def generate_prompt(
    task_description: str,
    model: str = "claude-opus-4-5",
    output_format: str = "JSON",
    edge_cases: str = "empty input, ambiguous cases",
    constraints: str = "respond in English only",
    input_format: str = "plain text",
) -> str:
    """Use an LLM to generate an optimized prompt for a given task."""
    import anthropic
    client = anthropic.Anthropic()
    
    response = client.messages.create(
        model=model,
        max_tokens=2000,
        messages=[{
            "role": "user",
            "content": META_PROMPT_GENERATOR.format(
                task_description=task_description,
                model=model,
                input_format=input_format,
                output_format=output_format,
                edge_cases=edge_cases,
                constraints=constraints,
            )
        }]
    )
    return response.content[0].text

7.6 Few-Shot Example Design

EXAMPLE DESIGN CHECKLIST:

[ ] 3-5 examples (more = diminishing returns + token cost)
[ ] Covers: simple case, edge case, complex case, negative case
[ ] Consistent format across all examples
[ ] Output format matches expected production output exactly
[ ] Examples do NOT appear in test set (data contamination)
[ ] Ordered: simple -> complex (progressive difficulty)

EXAMPLE TEMPLATE:

Input: {diverse_input}
Output: {correctly_formatted_output}

[Repeat for each example with blank line between]

Now apply to:
Input: {actual_input}
Output:

---

8. ML Integration for Agents

8.1 LLM Provider Abstraction Layer

from abc import ABC, abstractmethod
from dataclasses import dataclass
import time

import anthropic
import openai
from tenacity import retry, stop_after_attempt, wait_exponential

@dataclass
class LLMResponse:
    content: str
    model: str
    input_tokens: int
    output_tokens: int
    cost_usd: float
    latency_ms: float

# Cost per 1K tokens (update as pricing changes)
PRICING = {
    "claude-opus-4-5":     {"input": 0.015,    "output": 0.075},
    "claude-haiku-3":      {"input": 0.00025,  "output": 0.00125},
    "gpt-4o":              {"input": 0.005,    "output": 0.015},
    "gpt-4o-mini":         {"input": 0.00015,  "output": 0.0006},
}

class LLMProvider(ABC):
    @abstractmethod
    def complete(self, prompt: str, **kwargs) -> LLMResponse:
        pass
    
    def estimate_cost(self, prompt: str, output_tokens_estimate: int = 500) -> float:
        model = getattr(self, "model", "unknown")
        pricing = PRICING.get(model, {"input": 0.01, "output": 0.03})
        input_tokens = len(prompt.split()) * 1.3  # rough approximation
        return (input_tokens / 1000 * pricing["input"] +
                output_tokens_estimate / 1000 * pricing["output"])


class AnthropicProvider(LLMProvider):
    def __init__(self, model: str = "claude-opus-4-5"):
        self.client = anthropic.Anthropic()
        self.model  = model
    
    @retry(stop=stop_after_attempt(3), wait=wait_exponential(min=1, max=10))
    def complete(self, prompt: str, max_tokens: int = 2048, **kwargs) -> LLMResponse:
        start = time.time()
        response = self.client.messages.create(
            model=self.model,
            max_tokens=max_tokens,
            messages=[{"role": "user", "content": prompt}],
            **kwargs,
        )
        latency = (time.time() - start) * 1000
        pricing = PRICING.get(self.model, {"input": 0.015, "output": 0.075})
        cost = (response.usage.input_tokens  / 1000 * pricing["input"] +
                response.usage.output_tokens / 1000 * pricing["output"])
        return LLMResponse(
            content=response.content[0].text,
            model=self.model,
            input_tokens=response.usage.input_tokens,
            output_tokens=response.usage.output_tokens,
            cost_usd=cost,
            latency_ms=latency,
        )


class OpenAIProvider(LLMProvider):
    def __init__(self, model: str = "gpt-4o-mini"):
        self.client = openai.OpenAI()
        self.model  = model
    
    @retry(stop=stop_after_attempt(3), wait=wait_exponential(min=1, max=10))
    def complete(self, prompt: str, max_tokens: int = 2048, **kwargs) -> LLMResponse:
        start = time.time()
        response = self.client.chat.completions.create(
            model=self.model,
            max_tokens=max_tokens,
            messages=[{"role": "user", "content": prompt}],
            **kwargs,
        )
        latency = (time.time() - start) * 1000
        usage   = response.usage
        pricing = PRICING.get(self.model, {"input": 0.005, "output": 0.015})
        cost = (usage.prompt_tokens     / 1000 * pricing["input"] +
                usage.completion_tokens / 1000 * pricing["output"])
        return LLMResponse(
            content=response.choices[0].message.content,
            model=self.model,
            input_tokens=usage.prompt_tokens,
            output_tokens=usage.completion_tokens,
            cost_usd=cost,
            latency_ms=latency,
        )


class FallbackProvider(LLMProvider):
    """Try primary provider; fall back to secondary on failure."""
    
    def __init__(self, primary: LLMProvider, fallback: LLMProvider):
        self.primary  = primary
        self.fallback = fallback
    
    def complete(self, prompt: str, **kwargs) -> LLMResponse:
        try:
            return self.primary.complete(prompt, **kwargs)
        except Exception as e:
            print(f"Primary LLM failed ({e}), falling back...")
            return self.fallback.complete(prompt, **kwargs)

8.2 Model Deployment for Agent Serving

# Dockerfile for agent service
FROM python:3.11-slim

WORKDIR /app

# Install dependencies first (layer caching)
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

# Copy source
COPY src/ ./src/
COPY config/ ./config/

# Health check endpoint
HEALTHCHECK --interval=30s --timeout=5s --start-period=10s --retries=3 \
  CMD curl -f http://localhost:8080/health || exit 1

EXPOSE 8080
CMD ["uvicorn", "src.agent_server:app", "--host", "0.0.0.0", "--port", "8080", "--workers", "4"]

# src/agent_server.py — FastAPI agent serving wrapper
from fastapi import FastAPI, HTTPException, BackgroundTasks
from pydantic import BaseModel
import asyncio
import uuid

app = FastAPI(title="Agent Service", version="1.0.0")

class AgentRequest(BaseModel):
    query:        str
    session_id:   str | None = None
    max_steps:    int        = 10
    metadata:     dict       = {}

class AgentResponse(BaseModel):
    result:     str
    session_id: str
    steps_used: int
    cost_usd:   float
    latency_ms: float

@app.get("/health")
async def health():
    return {"status": "ok", "version": "1.0.0"}

@app.post("/agent/run", response_model=AgentResponse)
async def run_agent(request: AgentRequest):
    session_id = request.session_id or str(uuid.uuid4())
    # ... agent execution logic ...
    return AgentResponse(
        result="...",
        session_id=session_id,
        steps_used=0,
        cost_usd=0.0,
        latency_ms=0.0,
    )

@app.post("/agent/run-async")
async def run_agent_async(request: AgentRequest, background: BackgroundTasks):
    job_id = str(uuid.uuid4())
    background.add_task(_run_agent_task, job_id, request)
    return {"job_id": job_id, "status": "queued"}

@app.get("/agent/status/{job_id}")
async def get_status(job_id: str):
    # Fetch from job store (Redis, DB, etc.)
    return {"job_id": job_id, "status": "unknown"}

async def _run_agent_task(job_id: str, request: AgentRequest):
    # Run agent and persist result
    pass

8.3 Model Monitoring for Agents

# Drift detection for agent input distributions
from scipy.stats import ks_2samp
import numpy as np

def detect_input_drift(
    reference_inputs: list[str],
    current_inputs: list[str],
    threshold_p: float = 0.05,
) -> dict:
    """
    Detect distribution shift in agent input queries using
    token length distribution as a proxy metric.
    """
    ref_lengths = np.array([len(t.split()) for t in reference_inputs])
    cur_lengths = np.array([len(t.split()) for t in current_inputs])
    
    stat, p_value = ks_2samp(ref_lengths, cur_lengths)
    
    return {
        "drift_detected": p_value < threshold_p,
        "ks_statistic":   float(stat),
        "p_value":        float(p_value),
        "ref_mean_tokens": float(ref_lengths.mean()),
        "cur_mean_tokens": float(cur_lengths.mean()),
        "recommendation": (
            "Retrain or re-evaluate agent prompts — input distribution shifted significantly."
            if p_value < threshold_p else
            "No significant drift detected."
        ),
    }

# Alert thresholds for agent monitoring
AGENT_ALERT_THRESHOLDS = {
    "p95_latency_ms":     {"warning": 2000,  "critical": 5000},
    "error_rate_pct":     {"warning": 1.0,   "critical": 5.0},
    "cost_per_query_usd": {"warning": 0.05,  "critical": 0.20},
    "tool_failure_rate":  {"warning": 0.05,  "critical": 0.15},
    "token_overflow_rate":{"warning": 0.02,  "critical": 0.10},
}

8.4 Serving Strategy Selection

| Strategy | Latency | Throughput | Cost | Use Case | |---------|---------|-----------|------|----------| | FastAPI + Uvicorn | Low | Medium | Low | REST agent APIs, single-model | | Ray Serve | Medium | Very High | Medium | Multi-model pipelines, scaling | | Triton Inference | Very Low | Very High | Medium | GPU batch inference | | Serverless (Lambda/Cloud Run) | Cold-start medium | Auto-scale | Pay-per-use | Bursty agent tasks | | Streaming (SSE/WebSocket) | Apparent Low | Medium | Low | Conversational agents |

---

9. Skill & Capability Creation

9.1 SKILL.md Format Specification

Every Perplexity Computer skill must follow this exact format:

---
name: skill-name-with-hyphens
description: One or two sentences describing when to use this skill. Start with "Use when..." or describe the trigger conditions clearly.
license: MIT
metadata:
  author: your-username
  version: '1.0'
---

# Skill Title

Brief one-paragraph overview of the skill's purpose.

## When to Use
...

## Core Concepts
...

## Step-by-Step Process
...

## Examples
...

## Common Mistakes
...

9.2 Validation Pipeline

# Validate skill structure and frontmatter
cd /home/user/workspace && uvx --from skills-ref agentskills validate <skill-name>

# What the validator checks:
# OK First line is exactly ---
# OK YAML frontmatter present and parseable
# OK Required fields: name, description, license, metadata.author, metadata.version
# OK name matches directory name
# OK version is quoted string ('1.0' not 1.0)
# OK Skill directory exists at workspace/<skill-name>/SKILL.md
# OK No syntax errors in YAML block

9.3 Skill Quality Checklist

Before publishing any skill:

• [ ] First line of SKILL.md is exactly --- (three dashes, no spaces)
• [ ] All required YAML fields present (name, description, license, metadata.author, metadata.version)
• [ ] version is quoted: '1.0' not 1.0
• [ ] Description tells the agent WHEN to load the skill (trigger conditions)
• [ ] ## When to Use section with clear positive AND negative cases
• [ ] At least one working code example
• [ ] Common mistakes / red flags section
• [ ] Validation passes: uvx --from skills-ref agentskills validate <name>

9.4 Skill Packaging for Distribution

skill-directory/
+-- SKILL.md          <- Main skill file (required)
+-- README.md         <- Human-readable docs (optional)
+-- examples/
|   +-- basic.md      <- Annotated simple example
|   +-- advanced.md   <- Complex workflow example
+-- templates/
|   +-- prompt_template.md
|   +-- config.yaml
+-- scripts/
    +-- validate.py   <- Skill-specific validation helpers

9.5 Skill Trigger Design

The description field determines when the skill is loaded. Write it to trigger on the right signals:

# Bad — too vague, triggers on everything
description: Helps build things with AI.

# Bad — too narrow, misses many triggers
description: Use when the user types "build an MCP server".

# Good — triggers on intent, not exact phrasing
description: >
  Use when building AI agents, creating MCP servers, designing RAG systems,
  coordinating subagents, optimizing prompts, or architecting any AI-powered
  automation workflow. Covers agent design patterns, multi-agent orchestration,
  and production deployment.

---

10. Backend Infrastructure for Agents

10.1 Agent Memory Persistence with SQLite (CGI-bin)

Agents need persistent memory across sessions. The CGI-bin pattern lets agents store and retrieve state via HTTP endpoints without a dedicated backend server.

#!/usr/bin/env python3
# cgi-bin/agent_memory.py
# Agent memory store: conversations, tool results, learned facts

import json
import os
import sqlite3
import sys
from datetime import datetime

DB_PATH = "agent_memory.db"

def init_db(conn):
    conn.executescript("""
        CREATE TABLE IF NOT EXISTS sessions (
            session_id  TEXT PRIMARY KEY,
            created_at  TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
            metadata    TEXT DEFAULT '{}'
        );
        
        CREATE TABLE IF NOT EXISTS messages (
            id          INTEGER PRIMARY KEY AUTOINCREMENT,
            session_id  TEXT NOT NULL,
            role        TEXT NOT NULL CHECK(role IN ('user','assistant','tool','system')),
            content     TEXT NOT NULL,
            tool_name   TEXT,
            tool_input  TEXT,
            tool_result TEXT,
            created_at  TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
            FOREIGN KEY (session_id) REFERENCES sessions(session_id)
        );
        
        CREATE TABLE IF NOT EXISTS facts (
            id          INTEGER PRIMARY KEY AUTOINCREMENT,
            session_id  TEXT,
            key         TEXT NOT NULL,
            value       TEXT NOT NULL,
            confidence  REAL DEFAULT 1.0,
            created_at  TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
            expires_at  TIMESTAMP,
            UNIQUE(session_id, key)
        );
        
        CREATE INDEX IF NOT EXISTS idx_messages_session ON messages(session_id);
        CREATE INDEX IF NOT EXISTS idx_facts_key ON facts(key);
    """)
    conn.commit()

conn = sqlite3.connect(DB_PATH)
conn.row_factory = sqlite3.Row
init_db(conn)

method      = os.environ.get("REQUEST_METHOD", "GET")
query       = os.environ.get("QUERY_STRING", "")
path_info   = os.environ.get("PATH_INFO", "")

def respond(data, status=200):
    print(f"Status: {status}")
    print("Content-Type: application/json")
    print()
    print(json.dumps(data))

def parse_qs(qs):
    params = {}
    for part in qs.split("&"):
        if "=" in part:
            k, v = part.split("=", 1)
            params[k] = v
    return params

# -- Routes --------------------------------------------------------------------
if path_info == "/sessions" and method == "POST":
    body = json.loads(sys.stdin.read() or "{}")
    sid  = body.get("session_id") or f"sess_{datetime.utcnow().strftime('%Y%m%d_%H%M%S_%f')}"
    conn.execute("INSERT OR IGNORE INTO sessions (session_id, metadata) VALUES (?,?)",
                 [sid, json.dumps(body.get("metadata", {}))])
    conn.commit()
    respond({"session_id": sid}, 201)

elif path_info == "/messages" and method == "POST":
    body = json.loads(sys.stdin.read())
    conn.execute(
        "INSERT INTO messages (session_id, role, content, tool_name, tool_input, tool_result) "
        "VALUES (?,?,?,?,?,?)",
        [body["session_id"], body["role"], body["content"],
         body.get("tool_name"), body.get("tool_input"), body.get("tool_result")]
    )
    conn.commit()
    msg_id = conn.execute("SELECT last_insert_rowid()").fetchone()[0]
    respond({"id": msg_id}, 201)

elif path_info == "/messages" and method == "GET":
    params = parse_qs(query)
    sid    = params.get("session_id", "")
    limit  = int(params.get("limit", 50))
    rows   = conn.execute(
        "SELECT * FROM messages WHERE session_id=? ORDER BY created_at LIMIT ?",
        [sid, limit]
    ).fetchall()
    respond([dict(r) for r in rows])

elif path_info == "/facts" and method == "PUT":
    body = json.loads(sys.stdin.read())
    conn.execute(
        "INSERT OR REPLACE INTO facts (session_id, key, value, confidence) VALUES (?,?,?,?)",
        [body.get("session_id"), body["key"], json.dumps(body["value"]),
         body.get("confidence", 1.0)]
    )
    conn.commit()
    respond({"status": "ok"})

elif path_info == "/facts" and method == "GET":
    params = parse_qs(query)
    sid    = params.get("session_id", "")
    rows   = conn.execute(
        "SELECT key, value, confidence FROM facts WHERE session_id=? OR session_id IS NULL",
        [sid]
    ).fetchall()
    respond({r["key"]: {"value": json.loads(r["value"]), "confidence": r["confidence"]}
             for r in rows})

else:
    respond({"error": f"Unknown route: {method} {path_info}"}, 400)

10.2 Webhook Receiver for Agent Triggers

#!/usr/bin/env python3
# cgi-bin/webhook_receiver.py
# Receives external events and queues them for agent processing

import hashlib
import hmac
import json
import os
import sqlite3
import sys
import time

DB_PATH     = "webhook_events.db"
WEBHOOK_SECRET = os.environ.get("WEBHOOK_SECRET", "")

conn = sqlite3.connect(DB_PATH)
conn.execute("""
    CREATE TABLE IF NOT EXISTS events (
        id          INTEGER PRIMARY KEY AUTOINCREMENT,
        source      TEXT NOT NULL,
        event_type  TEXT NOT NULL,
        payload     TEXT NOT NULL,
        processed   INTEGER DEFAULT 0,
        received_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
    )
""")
conn.commit()

method = os.environ.get("REQUEST_METHOD", "GET")

def verify_signature(body: str, signature: str, secret: str) -> bool:
    """Verify HMAC-SHA256 webhook signature."""
    if not secret:
        return True  # Skip verification if no secret configured
    expected = "sha256=" + hmac.new(
        secret.encode(), body.encode(), hashlib.sha256
    ).hexdigest()
    return hmac.compare_digest(signature, expected)

if method == "POST":
    raw_body = sys.stdin.read()
    sig      = os.environ.get("HTTP_X_HUB_SIGNATURE_256", "")
    
    if not verify_signature(raw_body, sig, WEBHOOK_SECRET):
        print("Status: 401")
        print("Content-Type: application/json")
        print()
        print('{"error": "Invalid signature"}')
        sys.exit(0)
    
    body = json.loads(raw_body)
    conn.execute(
        "INSERT INTO events (source, event_type, payload) VALUES (?,?,?)",
        [body.get("source", "unknown"), body.get("type", "unknown"), raw_body]
    )
    conn.commit()
    
    print("Status: 202")
    print("Content-Type: application/json")
    print()
    print('{"status": "accepted"}')

elif method == "GET":
    # Dequeue unprocessed events for agent polling
    rows = conn.execute(
        "SELECT * FROM events WHERE processed=0 ORDER BY received_at LIMIT 50"
    ).fetchall()
    
    events = [
        {"id": r[0], "source": r[1], "event_type": r[2],
         "payload": json.loads(r[3]), "received_at": r[5]}
        for r in rows
    ]
    print("Content-Type: application/json")
    print()
    print(json.dumps(events))

10.3 Agent-to-Agent Communication via Message Bus

#!/usr/bin/env python3
# cgi-bin/message_bus.py
# Simple pub/sub message bus for multi-agent coordination

import json
import os
import sqlite3
import sys
import uuid

DB_PATH = "message_bus.db"

conn = sqlite3.connect(DB_PATH)
conn.executescript("""
    CREATE TABLE IF NOT EXISTS messages (
        id          TEXT PRIMARY KEY,
        from_agent  TEXT NOT NULL,
        to_agent    TEXT,           -- NULL = broadcast
        topic       TEXT NOT NULL,
        payload     TEXT NOT NULL,
        ack         INTEGER DEFAULT 0,
        created_at  TIMESTAMP DEFAULT CURRENT_TIMESTAMP
    );
    CREATE TABLE IF NOT EXISTS subscriptions (
        agent_id    TEXT NOT NULL,
        topic       TEXT NOT NULL,
        PRIMARY KEY (agent_id, topic)
    );
""")
conn.commit()

method    = os.environ.get("REQUEST_METHOD", "GET")
path_info = os.environ.get("PATH_INFO", "")
query     = os.environ.get("QUERY_STRING", "")

def respond(data, status=200):
    print(f"Status: {status}")
    print("Content-Type: application/json")
    print()
    print(json.dumps(data))

if path_info == "/publish" and method == "POST":
    body = json.loads(sys.stdin.read())
    msg_id = str(uuid.uuid4())
    conn.execute(
        "INSERT INTO messages (id, from_agent, to_agent, topic, payload) VALUES (?,?,?,?,?)",
        [msg_id, body["from_agent"], body.get("to_agent"),
         body["topic"], json.dumps(body["payload"])]
    )
    conn.commit()
    respond({"message_id": msg_id}, 201)

elif path_info == "/subscribe" and method == "POST":
    body = json.loads(sys.stdin.read())
    conn.execute(
        "INSERT OR IGNORE INTO subscriptions VALUES (?,?)",
        [body["agent_id"], body["topic"]]
    )
    conn.commit()
    respond({"status": "subscribed"})

elif path_info == "/poll" and method == "GET":
    params = dict(p.split("=") for p in query.split("&") if "=" in p)
    agent_id = params.get("agent_id", "")
    
    # Get messages for this agent (direct + subscribed topics)
    rows = conn.execute("""
        SELECT m.* FROM messages m
        LEFT JOIN subscriptions s ON s.agent_id=? AND s.topic=m.topic
        WHERE m.ack=0 AND (m.to_agent=? OR (m.to_agent IS NULL AND s.agent_id IS NOT NULL))
        ORDER BY m.created_at LIMIT 20
    """, [agent_id, agent_id]).fetchall()
    
    messages = [{"id": r[0], "from": r[1], "topic": r[3],
                 "payload": json.loads(r[4])} for r in rows]
    
    # Mark as acked
    if messages:
        ids = [m["id"] for m in messages]
        conn.execute(f"UPDATE messages SET ack=1 WHERE id IN ({','.join('?'*len(ids))})", ids)
        conn.commit()
    
    respond(messages)

10.4 JavaScript Client for Agent Backend

// Agent backend client — use __CGI_BIN__ placeholder in production
const CGI_BIN = "__CGI_BIN__";

class AgentMemoryClient {
  async createSession(metadata = {}) {
    const res = await fetch(`${CGI_BIN}/agent_memory.py/sessions`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({ metadata }),
    });
    return res.json();
  }

  async addMessage(sessionId, role, content, toolData = {}) {
    const res = await fetch(`${CGI_BIN}/agent_memory.py/messages`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({ session_id: sessionId, role, content, ...toolData }),
    });
    return res.json();
  }

  async getHistory(sessionId, limit = 50) {
    const res = await fetch(
      `${CGI_BIN}/agent_memory.py/messages?session_id=${sessionId}&limit=${limit}`
    );
    return res.json();
  }

  async setFact(sessionId, key, value, confidence = 1.0) {
    const res = await fetch(`${CGI_BIN}/agent_memory.py/facts`, {
      method: "PUT",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({ session_id: sessionId, key, value, confidence }),
    });
    return res.json();
  }

  async getFacts(sessionId) {
    const res = await fetch(
      `${CGI_BIN}/agent_memory.py/facts?session_id=${sessionId}`
    );
    return res.json();
  }
}

class MessageBusClient {
  async publish(fromAgent, topic, payload, toAgent = null) {
    const res = await fetch(`${CGI_BIN}/message_bus.py/publish`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({ from_agent: fromAgent, topic, payload, to_agent: toAgent }),
    });
    return res.json();
  }

  async subscribe(agentId, topic) {
    const res = await fetch(`${CGI_BIN}/message_bus.py/subscribe`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({ agent_id: agentId, topic }),
    });
    return res.json();
  }

  async poll(agentId) {
    const res = await fetch(`${CGI_BIN}/message_bus.py/poll?agent_id=${agentId}`);
    return res.json();
  }
}

---

11. Agent Deployment & Monitoring

11.1 Deployment Checklist

Before deploying any agent to production:

Infrastructure

• [ ] Health check endpoint responds: GET /health -> {"status": "ok"}
• [ ] Graceful shutdown handles in-flight requests
• [ ] Rate limiting configured (protect upstream APIs)
• [ ] Request timeouts set (prevent runaway agents)
• [ ] Retry logic with exponential backoff on transient failures
• [ ] Circuit breaker prevents cascade failures

Observability

• [ ] Structured logging (JSON) with correlation IDs
• [ ] Trace propagation across agent calls
• [ ] Metrics exported: request count, latency p50/p95/p99, error rate, token usage, cost
• [ ] Alerts configured for critical thresholds

Security

• [ ] API keys stored in environment variables, never in code
• [ ] Input sanitization before passing to LLM (prompt injection defense)
• [ ] Output filtering for PII / sensitive data
• [ ] Rate limits per user/tenant
• [ ] Audit log of all tool executions

Cost Controls

• [ ] Per-request token budget enforced
• [ ] Daily/monthly spend limits with alerting
• [ ] Model fallback chain configured (expensive -> cheap)

11.2 Structured Agent Logging

import json
import logging
import time
import uuid
from contextlib import contextmanager
from dataclasses import asdict, dataclass

@dataclass
class AgentTraceEvent:
    trace_id:    str
    span_id:     str
    event_type:  str   # "llm_call" | "tool_call" | "agent_start" | "agent_end"
    timestamp:   float
    agent_id:    str
    model:       str | None = None
    tool_name:   str | None = None
    input_tokens:  int = 0
    output_tokens: int = 0
    cost_usd:    float = 0.0
    latency_ms:  float = 0.0
    error:       str | None = None
    metadata:    dict | None = None

class AgentTracer:
    def __init__(self, agent_id: str):
        self.agent_id = agent_id
        self.logger   = logging.getLogger("agent.trace")
    
    def log(self, event: AgentTraceEvent):
        self.logger.info(json.dumps(asdict(event)))
    
    @contextmanager
    def trace_llm_call(self, trace_id: str, model: str):
        span_id = str(uuid.uuid4())[:8]
        start   = time.time()
        event   = AgentTraceEvent(
            trace_id=trace_id, span_id=span_id,
            event_type="llm_call", timestamp=start,
            agent_id=self.agent_id, model=model,
        )
        try:
            yield event
        except Exception as e:
            event.error = str(e)
            raise
        finally:
            event.latency_ms = (time.time() - start) * 1000
            self.log(event)
    
    @contextmanager
    def trace_tool_call(self, trace_id: str, tool_name: str):
        span_id = str(uuid.uuid4())[:8]
        start   = time.time()
        event   = AgentTraceEvent(
            trace_id=trace_id, span_id=span_id,
            event_type="tool_call", timestamp=start,
            agent_id=self.agent_id, tool_name=tool_name,
        )
        try:
            yield event
        except Exception as e:
            event.error = str(e)
            raise
        finally:
            event.latency_ms = (time.time() - start) * 1000
            self.log(event)

11.3 Agent Health Dashboard Schema

# Metrics to expose at /metrics (Prometheus-compatible)
AGENT_METRICS = """
# HELP agent_requests_total Total requests processed
# TYPE agent_requests_total counter
agent_requests_total{agent_id="{agent_id}",status="success"} {success_count}
agent_requests_total{agent_id="{agent_id}",status="error"} {error_count}

# HELP agent_latency_seconds Request latency
# TYPE agent_latency_seconds histogram
agent_latency_seconds_bucket{{agent_id="{agent_id}",le="0.5"}} {bucket_0_5}
agent_latency_seconds_bucket{{agent_id="{agent_id}",le="1.0"}} {bucket_1_0}
agent_latency_seconds_bucket{{agent_id="{agent_id}",le="5.0"}} {bucket_5_0}
agent_latency_seconds_bucket{{agent_id="{agent_id}",le="+Inf"}} {bucket_inf}

# HELP agent_tokens_total Total tokens consumed
# TYPE agent_tokens_total counter
agent_tokens_total{{agent_id="{agent_id}",type="input"}} {input_tokens}
agent_tokens_total{{agent_id="{agent_id}",type="output"}} {output_tokens}

# HELP agent_cost_usd_total Total cost in USD
# TYPE agent_cost_usd_total counter
agent_cost_usd_total{{agent_id="{agent_id}"}} {total_cost}

# HELP agent_tool_calls_total Tool calls by name
# TYPE agent_tool_calls_total counter
{tool_call_metrics}
"""

11.4 Scaling Strategies

| Scale Level | Approach | Infrastructure | |------------|---------|---------------| | Single user | Single process, local SQLite | Dev machine or single VM | | Small team (< 50 users) | Multi-worker Uvicorn, shared PostgreSQL | Single server, 4-8 CPU | | Medium (50-500 users) | Horizontal pod autoscaling, Redis cache | Kubernetes, load balancer | | Large (500+ users) | Async task queue (Celery/Arq), vector DB cluster | Multi-region, CDN | | Enterprise | Dedicated LLM endpoints, tenant isolation, SOC2 | Managed cloud |

---

12. Unique Perplexity Computer Capabilities

12.1 400+ Service Integrations

Perplexity Computer exposes connections to 400+ external services via the list_external_tools / call_external_tool pattern. This makes it uniquely suited for building integration-heavy agents without writing custom connectors.

Discovery pattern:

1. list_external_tools(queries=["github", "repo"])
   -> Returns connected tools with source_id and status

2. describe_external_tools(source_id="github", tool_names=["create_issue"])
   -> Returns full JSON schema for tool inputs

3. call_external_tool(tool_name="create_issue", source_id="github", arguments={...})
   -> Executes against live service

Key integration categories available: | Category | Example Services | |---------|-----------------| | Communication | Gmail, Slack, Teams, Discord, Outlook | | Project management | GitHub, Jira, Linear, Asana, Notion | | CRM/Sales | Salesforce, HubSpot, Pipedrive | | Data/Analytics | Google Sheets, Airtable, BigQuery | | Storage | Google Drive, Dropbox, S3 | | Calendar | Google Calendar, Outlook Calendar | | Social | Twitter/X, LinkedIn | | Payments | Stripe | | Databases | PostgreSQL, MySQL, MongoDB |

12.2 Building Integration-Driven Agents

An agent in Perplexity Computer that uses external integrations:

# Pattern: Research -> Enrich -> Store -> Notify
# Example: Competitor monitoring agent

COMPETITOR_AGENT_FLOW = """
1. search_web(queries=["[competitor] product update", "[competitor] pricing 2026"])
   -> Gather news about target competitor

2. search_social(query="from:[competitor_twitter] -is:retweet", only_recent=True)
   -> Collect recent social posts

3. call_external_tool(source_id="notion_mcp", tool_name="create_page", arguments={
       "title": f"Competitor Intel: {date}",
       "content": compiled_findings
   })
   -> Store findings in Notion

4. call_external_tool(source_id="gmail", tool_name="send_email", arguments={
       "to": "[email protected]",
       "subject": f"Weekly Competitor Brief: {date}",
       "body": executive_summary
   })
   -> Notify stakeholders
"""

12.3 Scheduled Monitoring Agents

Perplexity Computer supports recurring tasks that trigger agents on a schedule. Use this for:

• Price monitoring: scrape competitor pricing daily, alert on changes
• Brand mention tracking: search social + news, weekly digest
• Performance monitoring: check site metrics, alert on regressions
• Data drift detection: compare model input distributions, trigger retraining

Scheduled agent pattern:

MONITORING_AGENT_PROMPT = """
You are a scheduled monitoring agent running at: {timestamp}
Your monitoring target: {target_description}

## Steps
1. Collect current data from: {data_sources}
2. Compare to baseline stored at: {baseline_reference}
3. Calculate delta metrics
4. If any metric exceeds threshold: {alert_thresholds}
   -> Call alert tool with details
5. Update baseline with today's snapshot
6. Output summary report

## Output
Provide a structured report:
- Status: NORMAL | WARNING | CRITICAL
- Changes detected: [list]
- Metrics compared to baseline: [table]
- Actions taken: [list]
- Next check: {next_check_time}
"""

12.4 Research-Backed Agent Responses

Perplexity Computer's search_web, search_vertical, and fetch_url tools give agents access to current, real-world information that LLMs alone cannot provide.

Research agent pattern:

RESEARCH_AGENT_STEPS = [
    # Step 1: Broad search
    "search_web(queries=[query, related_query_1, related_query_2])",
    
    # Step 2: Deep dive on key sources
    "fetch_url(url=top_result_url, prompt='Extract key facts and data')",
    
    # Step 3: Academic grounding (for technical claims)
    "search_vertical(vertical='academic', query=technical_query)",
    
    # Step 4: Visual evidence
    "search_vertical(vertical='image', query=visual_query)",
    
    # Step 5: Synthesize with citations
    "write synthesis with inline markdown citations linking to sources",
]

12.5 Live Deployment for Agent Interfaces

Agents built on Perplexity Computer can be deployed as live, publicly accessible web applications using the deploy_website workflow.

Agent Interface Deployment Stack:

Frontend (HTML/JS) --> deploy_website(project_path)
     |
     | fetch(__CGI_BIN__/agent_memory.py/...)
     |
Backend (CGI-bin Python) --> auto-deployed with frontend
     |
     | sqlite3 / flat files
     |
Persistent Storage --> lives in project directory

Result: A live URL with full agent interface accessible from anywhere.
No separate server provisioning needed.

Full-stack agent UI deployment checklist:

• [ ] index.html contains the agent chat interface
• [ ] cgi-bin/agent_memory.py handles session/message persistence
• [ ] cgi-bin/agent_api.py proxies LLM calls (keeps API key server-side)
• [ ] All CGI scripts are marked executable: chmod +x cgi-bin/*.py
• [ ] Client JavaScript uses __CGI_BIN__ as base URL (replaced at deploy time)
• [ ] CORS is handled by the CGI proxy layer (no browser CORS issues)
• [ ] Deployed with deploy_website tool

---

Appendix A: Quick Reference — When to Use Which Pattern

| Situation | Use This | |-----------|----------| | Need to build an agent from scratch | §2.2 Architecture Patterns — pick ReAct, Plan-Execute, or Reflexion | | Need to integrate an external API as an agent tool | §3 MCP Server Development | | Need the agent to answer from private documents | §4 RAG System Construction | | Have multiple tasks that can run independently | §5.5 Parallel Agent Dispatch (dispatching-parallel-agents) | | Have sequential tasks with quality gates | §5.1 Subagent-Driven Development | | Have a written plan to implement | §6.3 Batch Execution with Checkpoints (executing-plans) | | Prompt producing inconsistent or low-quality output | §7 Prompt Engineering & Optimization | | Need to swap LLM providers or add cost controls | §8.1 LLM Provider Abstraction Layer | | Need to package a workflow as a reusable skill | §9.1 SKILL.md Format Specification | | Need persistent agent state across sessions | §10.1 Agent Memory Persistence (SQLite/CGI-bin) | | Need agents to communicate with each other | §10.3 Agent Message Bus | | Deploying an agent to production | §11 Agent Deployment & Monitoring | | Using Perplexity Computer for agent automation | §12 Unique Perplexity Computer Capabilities |

---

Appendix B: Architecture Decision Records (ADR)

ADR-001: ReAct vs. Plan-Execute

Context: Choosing architecture for a new agent. Decision: Use ReAct for open-ended tasks; Plan-Execute for structured workflows. Rationale: ReAct handles unknowns gracefully. Plan-Execute gives auditability and checkpointing needed for long-running structured tasks.

ADR-002: Vector DB Selection for RAG

Context: Choosing vector database for production RAG. Decision: Pinecone for managed production; Chroma for local development. Rationale: Pinecone eliminates ops burden at production scale. Chroma has zero-setup for dev/test.

ADR-003: MCP Language Choice

Context: TypeScript vs. Python for MCP servers. Decision: TypeScript as default; Python (FastMCP) when team is Python-only. Rationale: TypeScript SDK has broader client compatibility. Static typing catches tool schema errors at compile time.

ADR-004: Subagent Review Order

Context: Should spec review or code quality review happen first? Decision: Spec compliance ALWAYS before code quality. Rationale: Code quality review on spec-non-compliant code wastes review cycles. Fixing spec gaps may invalidate quality feedback.

ADR-005: Agent Memory Architecture

Context: How should agents persist state in Perplexity Computer? Decision: SQLite via CGI-bin for development; PostgreSQL/Redis for production. Rationale: CGI-bin SQLite requires zero infrastructure, deploys with the frontend. Swap to PostgreSQL when multi-instance or high-traffic.

---

Appendix C: Evaluation Question Patterns for MCP Servers

When creating the 10 evaluation questions required by Phase 4 of MCP development:

GOOD EVALUATION QUESTIONS:

OK Multi-step: requires 3+ tool calls to answer
OK Read-only: only non-destructive operations
OK Verifiable: single correct answer, checkable by string comparison
OK Realistic: something a real user would ask
OK Stable: answer won't change over time
OK Independent: not dependent on other questions

BAD EVALUATION QUESTIONS:

NO Single-step: answerable with one tool call
NO Write operations: creates, updates, or deletes data
NO Ambiguous: multiple valid answers
NO Unstable: answer changes (e.g., "latest version")
NO Dependent: requires previous question's state

Example good evaluation question:

<qa_pair>
  <question>
    Find all issues labeled "bug" and "high-priority" in the repository.
    What is the title of the oldest open one, and which user has been assigned
    the most bugs in that same repository?
  </question>
  <answer>Oldest: "Memory leak in session handler" | Most assigned: alice (7 bugs)</answer>
</qa_pair>

---

Appendix D: Common Failure Modes & Fixes

| Failure Mode | Symptoms | Fix | |-------------|---------|-----| | Context window overflow | Agent truncates history, loses tool results | Implement sliding window memory, summarize old turns | | Tool call hallucination | Agent calls nonexistent tools | Enumerate tools explicitly in system prompt; use native tool calling API | | Prompt injection | User input overrides agent instructions | Wrap user input in XML tags: <user_input>{input}</user_input> | | Infinite ReAct loop | Agent never reaches Final Answer | Add explicit iteration counter; add "if unsure after N steps, state limitations" | | Parallel agent conflicts | Two agents edit same file | Map files to agents before dispatch; check check_result_conflicts() | | RAG hallucination | Agent answers outside retrieved context | Add "Only answer from context. Say 'I don't know' if not in context." | | Spec creep | Implementer adds unasked-for features | Spec reviewer checks for EXTRA features; reject scope creep explicitly | | Cost overrun | Agent exceeds budget | Set max_tokens per request; add total-session token budget | | Stale memory | Agent uses outdated facts | Add TTL to fact store; validate facts against current context | | MCP tool discovery failure | Agent can't find right tool | Use consistent naming: service_verb_noun; add synonyms to description |