Swarm Workflow Protocol

• 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 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

Standard Handoff Format

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

File Locations

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)