Multi-Agent Harness

What You Get

The Multi-Agent Harness (MAH) is a TypeScript CLI and Next.js dashboard that orchestrates AI agents through a structured development cycle. Instead of one agent writing and evaluating its own code (AI grading its own homework), MAH runs dedicated agents for planning, development, QA, and recovery — each with clear contracts and specialized skills.

The result: automated development sprints that pass QA at 94% with an average cost of $0.03 each.

How It Works

1. Planning — You describe what you want. An Opus-class model breaks it into focused sprints and assigns the right agent.

2. Contract negotiation — The dev agent proposes a definition of done. The QA agent tightens it. Both agree before a single line of code is written.

3. Execution — Dev agent builds. QA agent evaluates against the contract. Failures route back to dev automatically — up to 3 rounds.

4. Recovery — Crashed sprints resume from transcript. No wasted work on restarts.

5. Dashboard — Live sprint execution, kanban board, cost tracking, sprint builder, and agent config — all in one UI.

The system is self-hosting: it built its own dashboard over four sprints, for under $1 total.

AI Infrastructure

Multi-Agent Harness

19 automated dev sprints over one weekend. 94% QA pass rate. $0.03/sprint average cost. The system built its own dashboard.

94% pass rateRead case study →

MAH Starter — $1,500

Our proven 5-agent setup, cloned and configured for your projects.

What's included

• 5 specialized agents — frontend dev, backend dev, QA evaluator, research analyst, content writer
• Pre-configured quality tiers — each agent has standards, anti-patterns, and grading criteria tuned for your stack
• Full MAH pipeline — planning, contract negotiation, execution, auto-recovery
• Next.js dashboard — live sprint view, kanban board, cost tracking, sprint builder
• 1 week of setup — we configure, test, and walk you through the system
• Documentation — runbooks for your team

Timeline

Running your first sprint within a week of kickoff. Most teams are self-sufficient within two weeks.

Best for

• Engineering teams building software with AI tooling today
• Teams that want repeatable, QA'd AI development without babysitting agents
• Anyone already using Claude Code, Codex, or OpenClaw

MAH Custom — $4,000

Custom agent roster designed for your specific workflow and codebase.

What's included

Everything in MAH Starter, plus:

• Codebase analysis — we study your architecture, conventions, and patterns before writing a single agent prompt
• Custom agent roster — agents tailored to your domain (e.g. mobile dev, data pipeline, API integrations, compliance review)
• Custom evaluation criteria — QA grading calibrated to your quality standards, not generic ones
• Specialized skills — agents get domain knowledge loaded for your stack and tools
• Full harness setup — pipeline tuned to your deployment target (GitHub Actions, CI/CD, local, cloud)
• You own everything — no recurring fees, no vendor lock-in, full source code

Timeline

2–3 weeks from kickoff to first automated sprint.

Best for

• Teams with specific stacks or domain requirements
• Organizations that need QA criteria aligned to internal standards
• Codebases where generic agents produce too much noise

Compatible Platforms

MAH works with any agent runtime:

• Claude Code — what we use for dev agents
• OpenClaw — agent orchestration with skills, memory, and multi-channel support
• NemoClaw — NVIDIA's enterprise policy engine for sandboxed execution
• Codex — OpenAI's agent platform
• Custom agents — anything that takes a prompt and returns output

What's Not Included

• Ongoing agent execution costs (billed to your API account — typically $0.03–$0.10/sprint)
• Custom integrations beyond the standard pipeline (available as add-ons)
• Managed hosting (ask about options if you need it)

Both tiers include a week of setup, a walkthrough of the system, and documentation for your team.

Interested? Book a call or email greg@gtalabs.com.