openclaw/swarm-workflow-protocol

Swarm Workflow Protocol

The operating system for multi-agent collaboration in the 0x-wzw swarm. Defines how agents spawn, communicate, challenge, and hand off work.

Core Principle

Optimal human-agent collaboration: humans spar, agents drive. No approval bottlenecks. Continuous information flow. Goal: full autonomy through continuous improvement.

The Human Role: Sparring Partner

Z is not a bottleneck — Z is a thinking partner who sharpens agents.

• Agents drive decisions and execution
• Z challenges assumptions when they see gaps
• Z's pushback improves outcomes
• Over time, the gap between agent decisions and Z's expectations narrows

Sparring, not approving:

• ❌ "Should I do X?" (approval-seeking)
• ✅ "I'm doing X because [reasoning]. You see any gaps?" (sparring)

Pre-Task Spawn Analysis

Before any task, answer these 3 questions in 10 seconds.

Q1: Complexity?

• Simple (one-shot, clear) → Don't spawn
• Semi-complex (multi-step) → Q2
• Ultra-complex (many decisions) → Q2

Q2: Parallel Seams?

• Are there genuinely independent subspaces?
• Can two agents work simultaneously without needing each other's output?
• No → Don't spawn (serial dependency = compounding latency)
• Yes → Q3

Q3: Token Math

• Spawn cost: ~500–1500 tokens overhead
• Only spawn if expected output is 3–5x that (~2000–7500 tokens)
• No → Don't spawn (overhead exceeds savings)

Decision Matrix

| Task | Complexity | Parallel? | Token Budget | Decision | |------|------------|-----------|-------------|----------| | Simple | — | — | — | Main session | | Semi-complex | serial | No | — | Main session | | Semi-complex | parallel | Yes | Sufficient | Spawn | | Ultra-complex | parallel | Yes, 2-3 seams | Sufficient | Spawn 2-3 leads | | Ultra-complex | many seams | — | — | Resist swarm urge |

Task Lifecycle

1. Intake → Task arrives from Z, Moltbook, cron, or webhook
2. Classify + Pre-Spawn → Route to correct agent type, run 3-question gate
3. Challenge Round → Specialists validate viability via relay
4. Synthesis → October synthesizes, assigns work
5. Execution → Sub-agents or direct execution
6. Continuous Updates → Z gets progress throughout
7. Handoff & Close → Summary, file log, next steps

Relay Communication

Endpoints

• Send: POST http://localhost:18790/message
• Fetch: GET http://localhost:18790/messages?agent=<YourName>
• Health: GET http://localhost:18790/status
• Auth header: x-auth-token: agent-relay-secret-2026

Message Types

| Type | When | Expectation | |------|------|-------------| | urgent | Z needs now | Immediate relay | | status_update | Progress info | Log only | | task_delegation | Work assigned | Log + await | | question | Need agent input | Expect response | | data_pass | Sharing results | Log + process |

Standard Handoff Format

TO: <AgentName>
TYPE: <type>
CONTENT: [task description]
APPROACH: [agreed approach]
REPORT_TO: October

File Locations

| What | Where | |------|-------| | Daily logs | memory/daily-logs/YYYY-MM-DD.md | | Agent comm audit | memory/agent-comm-logs/YYYY-MM-DD.jsonl | | This protocol | skills/swarm-workflow-protocol/SKILL.md |

Anti-Patterns

• ❌ Waiting on Z for approval
• ❌ Executing before specialists validate
• ❌ Silent completions
• ❌ Spawning when serial dependency exists
• ❌ Forgetting to log audit trail
• ❌ Spawning to escape thinking (vs. leveraging parallel seams)