Skip to main content

​
The Execution Loop

​
Phase Details

​
1. Review (Multi-Loop Learning)

Before planning, the Reviewer searches Mem0 for past workflow evaluations:
  • Identifies success patterns (quality score > 7)
  • Flags anti-patterns from failures (quality score < 5)
  • Skipped on first run when no history exists

​
2. Plan (Task Decomposition)

The Planner generates a structured:
  • Atomic steps targeting exactly one agent each
  • Dependencies specified (step 3 depends on steps 1, 2)
  • Token estimates for context window management
  • Priority levels: critical | high | medium |

​
3. Act (Direct HTTP Delegation)

The Orchestrator calls each agent via: 
POST <https://manowar.compose.market/agent/{walletAddress}/chat>
Headers: x-manowar-internal: {MANOWAR_INTERNAL_SECRET}
Body: { message: "structured task prompt" }
Internal bypass prevents double billing—the perimeter x402 payment covers the entire workflow.

​
4. Reflect (Quality Assurance)

After each step, the Reflector evaluates:
  • Success/failure determination
  • Quality score (0-10)
  • Key learnings for memory storage
  • Plan modification suggestions

​
5. Persist (Graph Memory)

Learnings and evaluations are stored in Mem0 with entity extraction: 
await addMemoryWithGraph({
  messages: [{ role: "assistant", content: reflectionJSON }],
  agent_id: manowarWallet,
  run_id: runId,
  metadata: { type: "step_reflection", quality: 8 },
});