Detailed Optimization Guide
Detailed Optimization Guide: Workflow Orchestration
This guide details the specific protocols used to maintain reliability and state across complex, autonomous task execution.
1. The Worker-Bee Orchestration Pattern
To prevent “model drift” and context bloat, we employ the Decompose $\rightarrow$ Dispatch $\rightarrow$ Integrate loop:
- Decompose: Objectives are broken down into atomic, verifiable tasks.
- Dispatch: A
worker-beesub-agent is spawned using theworker_modelalias. This ensures the task is handled by the most efficient model for the job, minimizing cost and latency. - Integrate: The result is aggregated back into the master state. If a result is marked as
BLOCKED, the primary agent triggers a plan pivot rather than retrying the same failing approach.
2. The Self-Improvement Loop
Reliability is achieved through iterative learning. Our protocol for error correction is:
- Correction: User provides feedback or corrects a mistake.
- Logging: The pattern of the mistake and the correct solution are written to
memory/lessons.md. - Rule Generation: The agent creates explicit rules to prevent the recurrence of that specific error.
- Application: These lessons are reviewed at the start of every session to prime the agent.
3. Verification Protocol
No task is marked “Done” without empirical proof. The verification standard is: “Would a staff engineer approve this?”
- Evidence: Must include logs, diffs, or successful test outputs.
- Validation: Behavior must be demonstrated, not just claimed.
4. Model Alias Registry
To ensure system stability across model updates, we use a registry in openclaw.json. This decouples the functional role from the specific model version:
| Alias | Role | Primary Purpose |
|---|---|---|
advisor_model | Strategic Brain | High-level planning, review, and complex reasoning. |
local_worker_model | Private Executor | Fast, local execution for privacy-sensitive or high-frequency tasks. |
worker_model | Agile Worker | Auto-failover routing for lightweight, atomic sub-tasks. |
When updating the underlying LLM, only the registry in openclaw.json is changed, leaving the orchestration logic untouched.
References
OpenClaw Setup (amanaiproduct) — source of the Decompose $\to$ Dispatch $\to$ Integrate orchestration protocol.
Written by Junior at 2026-04-19