beam-ai-team/create-beam-agent

Create Beam Agent

Deploy a Beam agent from a YAML description using scripts/create_agent_from_prompt.py. The skill translates the YAML into a typed JSON spec and calls the Beam complete graph API.

---

Prerequisites

• BEAM_API_KEY and BEAM_WORKSPACE_ID set in .env
• requests Python package installed (pip3 install requests)
• Script at scripts/create_agent_from_prompt.py (in project root)

---

Key Rules (must follow every time)

1. Entry node is always bare — no toolConfiguration, no params, no prompt. Objective = "Entry Node". If the YAML marks a processing node as is_entry: true, create a separate bare entry node and connect it to that processing node.

2. toolFunctionName is auto-generated by the script as GPTAction_Custom_{CamelCaseName} from the tool_name. Never set it manually in the spec.

3. Linked params use linked_node + linked_param in the spec (not linked_node_id as YAML may use). linked_node = the key of the source node in the spec.

4. Spec node keys = snake-case IDs matching the YAML id fields. The script generates all UUIDs — never put UUIDs in the spec.

5. on_error: "CONTINUE" only for non-critical nodes (e.g. Slack notifications). All others use "STOP".

---

Spec Format

{
  "agentName": "string",
  "agentDescription": "string",
  "personality": "string",
  "restrictions": "string",
  "prompts": ["example prompt 1", "example prompt 2"],
  "nodes": [
    {
      "key": "entry",
      "objective": "Entry Node",
      "is_entry": true,
      "x": 250, "y": 0,
      "edges": [{ "target": "first-processing-node", "name": "", "condition": "" }]
    },
    {
      "key": "node-key",
      "name": "Tool Display Name",
      "objective": "What this node does",
      "is_entry": false,
      "x": 250, "y": 200,
      "model": "BEDROCK_CLAUDE_SONNET_4",
      "tool_name": "Tool Display Name",
      "tool_description": "One-line description",
      "prompt": "Full LLM instruction prompt",
      "on_error": "STOP",
      "enable_retry": false,
      "retry_count": 1,
      "retry_wait_ms": 1000,
      "fallback_models": null,
      "evaluation_criteria": [],
      "input_params": [
        {
          "name": "param_name",
          "description": "what this param is",
          "type": "string|object|number|boolean",
          "is_array": false,
          "fill_type": "static|linked|user_fill|ai_fill",
          "static_value": null,
          "linked_node": null,
          "linked_param": null,
          "output_example": null,
          "required": true,
          "position": 0
        }
      ],
      "output_params": [
        {
          "name": "param_name",
          "description": "what this output contains",
          "type": "string|object|number|boolean",
          "is_array": false,
          "output_example": null,
          "position": 0
        }
      ],
      "edges": [
        { "target": "next-node-key", "name": "Edge label", "condition": "" }
      ]
    }
  ]
}

Positioning

• Linear flow: entry at y=0, processing nodes at y=200, x increases by 300 per step
• Conditional branches: fan out with y offsets (e.g. y=0 / y=300 / y=600)

fill_type values

| Value | When to use | |---|---| | static | Fixed hardcoded value — set static_value | | linked | Flows from a parent node's output — set linked_node + linked_param | | user_fill | User provides at runtime (first processing node inputs) | | ai_fill | AI extracts from conversation context automatically |

Available models

• BEDROCK_CLAUDE_SONNET_4 — default, fast, most tasks
• BEDROCK_CLAUDE_OPUS_4_5 — complex generation/reasoning

Conditional edges

• Unconditional: "condition": ""
• Conditional: "condition": "sum is odd" / "condition": "status is approved" etc.

---

Workflow

Step 1: Parse the YAML

Read the YAML agent spec and extract:

• Agent-level: name, description, personality, restrictions, prompts
• For each node: id (→ key), objective, x/y, on_error, retry settings, tool.*, edges
• For each input param: name, description, type, fill_type, static_value, linked_node_id (→ linked_node), linked_param, required, position, is_array
• For each output param: name, description, type, is_array, position

Step 2: Build the Spec JSON

Apply the rules:

• Prepend a bare entry node that connects to the first processing node
• Set is_entry: false on all processing nodes (even if YAML said is_entry: true)
• Map YAML linked_node_id → spec linked_node (uses the key / YAML id value)
• Set positions: entry at y=0, processing nodes at y=200, x increments by 300

Step 3: Write Spec to Temp File

Write the spec JSON to /tmp/beam_agent_spec.json.

Step 4: Dry Run

python3 scripts/create_agent_from_prompt.py --spec-file /tmp/beam_agent_spec.json --dry-run 2>&1 | head -30

Verify the node summary output looks correct (node names, input/output counts, linked params).

Step 5: Create Agent

python3 scripts/create_agent_from_prompt.py --spec-file /tmp/beam_agent_spec.json

Report the Agent ID and Draft Graph ID to the user.

---

Example: 4-node linear agent

Given YAML with nodes: extract → transform → generate-report → notify

Spec nodes:

entry (bare, y=0) → extract-yardi-data (y=200, x=250) → transform-validate (y=200, x=550) → generate-report (y=200, x=850) → notify-team (y=200, x=1150)

---

Error Handling

| Error | Cause | Fix | |---|---|---| | Linked param 'node.param' not found | linked_node or linked_param typo | Check the source node's key and output param name match exactly | | API Error 400 | Malformed payload | Run --dry-run, inspect JSON for missing fields | | API Error 401 | Bad API key | Check BEAM_API_KEY in .env | | Agent created but nodes empty | Wrong toolFunctionName prefix | Script auto-handles — ensure tool_name is set on every non-entry node |

---

Script Reference

create_agent_from_prompt.py

Usage:
  python3 scripts/create_agent_from_prompt.py --spec-file FILE
  python3 scripts/create_agent_from_prompt.py --spec-file FILE --dry-run
  echo '<json>' | python3 scripts/create_agent_from_prompt.py

What it does:
  - Pre-generates UUIDs for all nodes, tool configs, output params
  - Resolves linked params by UUID lookup
  - Builds originalTool block (required by API)
  - POSTs to /agent-graphs/complete (draft status)
  - Returns Agent ID, Draft Graph ID, Active Graph ID