Deterministic Multi-Agent Dev Pipeline

Overview

A fully deterministic code → review → test pipeline using OpenClaw's Lobster workflow engine with newly-contributed loop support. This architecture separates creative work (LLMs) from orchestration (code).

The Innovation

Contributor: Gustavo Gondim (DEV Community)
Date: February 23, 2026
Achievement: Contributed sub-workflow steps with loop support to Lobster (PR #20)

Architecture

Agent Roles

Agent	Model	Access	Responsibility
Programmer	Claude Opus	Full tools	Write code
Reviewer	Claude Sonnet	Read-only	Review code
Tester	Various	Test runners	Run tests

Key Technical Breakthroughs

Sub-workflow loops - Code→review cycle loops up to 3 times until approved
Deterministic routing - Via agentToAgent and sessions_send
Session key addressing - Pattern: pipeline:<project>:<role>
YAML state machine - Controls flow, not LLM

Workflow

# dev-pipeline.lobster
name: dev-pipeline
steps:
  - id: code-review-loop
    lobster: ./code-review.lobster
    loop:
      maxIterations: 3
      condition: 'approved == false'
  
  - id: test
    command: notify_tester
    condition: approved == true

Key Insight

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

Implementation Details

Session Addressing

pipeline:project-a:programmer
pipeline:project-a:reviewer
pipeline:project-b:tester

Communication Primitives

agentToAgent - Native peer messaging
sessions_send - Addressable session messaging
Webhooks with session routing

Results

100% autonomous pipeline execution
Deterministic flow control
Cost-optimized model routing
GitHub Copilot wrote the implementation

Use Cases

Automated code review pipelines
Multi-stage testing workflows
Continuous integration automation
Quality assurance gates

References

DEV Community Article (February 23, 2026)
Lobster PR #20
GitHub Copilot coding agent