Deterministic multi-agent dev pipeline with Lobster workflows
Deterministic Multi-Agent Dev Pipeline with Lobster Workflows in OpenClaw
A developer built a deterministic multi-agent development pipeline inside OpenClaw using the Lobster workflow engine. Three isolated agents (programmer, reviewer, tester) work autonomously with YAML-driven orchestration instead of letting an LLM decide the flow. He contributed the missing loop support back to the open-source project.
YAML-Driven Orchestration: Programmer, Reviewer, Tester with Schema Validation
After two months of exploring alternatives (Copilot agent sessions, building Protoagent, evaluating Ralph Orchestrator), the developer found that Lobster was the right foundation but lacked loops. His solution:
- Deterministic orchestration: YAML workflows control the flow, not LLMs. The state machine decides when the reviewer runs, not a prompt
- Three specialized agents: Programmer writes code, Reviewer checks it (max 3 iterations), Tester validates
- Parallel execution: 4 projects x 3 roles = up to 12 concurrent agent sessions
- Structured JSON validation: LLM output is validated against schemas before it triggers the next step
- Sub-lobsters: Nested workflow feature he contributed to OpenClaw/Lobster (PR #20)
Each agent gets its own tools, memory, identity, and workspace. Communication happens via Telegram.
Sub-lobsters Feature Contributed to OpenClaw Open Source (PR #20)
Contributed the Sub-lobsters feature to the OpenClaw/Lobster open-source project. Built a working code-to-review-to-test pipeline with deterministic flow control. GitHub Copilot coding agent wrote 100% of the implementation.
Built by Ggondim
Ggondim, published on dev.to with full technical architecture details.