Deterministic Multi-Agent Dev Pipeline

A fully deterministic code → review → test pipeline using OpenClaw's Lobster workflow engine with newly-contributed loop support.

The Problem with LLM Orchestration

Traditional multi-agent systems rely on LLMs to decide what happens next. This leads to:

• Non-deterministic behavior
• Difficult debugging
• Unpredictable costs
• Hard to reproduce issues

The Solution: YAML State Machine

workflow:
  steps:
    - name: code
      agent: programmer
      output: code_result
    - name: review
      agent: reviewer
      input: code_result
      output: review_result
    - name: test
      agent: tester
      input: review_result
      loop: until_pass

Key Innovation

• Deterministic routing - YAML controls flow, not LLM
• Agent-to-agent messaging - sessions_send for reliable communication
• Session keys as addressing - pipeline:<project>:<role>

Workflow

1. Programmer Agent - Writes code based on requirements
2. Reviewer Agent - Reviews code for quality and bugs
3. Tester Agent - Runs tests, reports results
4. Loop until pass - Cycles back if tests fail

Results

• Predictable execution
• Reproducible builds
• Lower token costs
• Easier debugging

Source

Gustavo Gondim on DEV Community, February 23, 2026.

Key Quote: "Don't orchestrate with LLMs. Use them for creative work, use code for plumbing."