We are open sourcing the starter kit we use to bootstrap production agent harnesses - Planner/Executor/Critic scaffolding, scoped tool permissions, and structured handoffs. Here is what is inside, how to use it, and where to grab the code on GitHub.
Why We Are Open Sourcing This
Every production agent system we ship at TunerLabs starts from the same skeleton: a Planner that decomposes work, an Executor with scoped write access, and a Critic that grades the result before anything merges. We have rebuilt that skeleton enough times to know it should live in one place, versioned and shared, instead of being copied between projects.
So we packaged it. The Agent Harness Starter Kit is the opinionated, production-ready foundation we use to bootstrap new agent work - and it is now open source under a permissive license.
> Want help adapting this to your stack? Book a free 30-minute architecture review and we will walk through your harness with one of our engineers.
What Is Inside
The kit ships with three things, and nothing you do not need:
- Planner/Executor/Critic scaffolding - the three-role loop wired up with typed message handoffs between stages, so each agent receives exactly the context it should and nothing more.
- Scoped tool permissions - read-only defaults, with write and shell access added per-role through explicit allow lists. The Planner and Critic cannot mutate anything; only the Executor can.
- Structured handoffs - schema-validated JSON blocks at every agent boundary, so format drift fails loudly at the seam instead of silently three stages later.
That is the whole surface. The kit is intentionally small. It is a foundation to build on, not a framework to fight.
Get the Code
Everything lives on GitHub. Clone the main repository, browse the examples, or read the contributing guide if you want to send a patch:
How To Use It
Clone the repository, drop your task spec into the Planner stage, and wire your tools into the Executor's allow list. The Critic comes pre-seeded with a small set of invariants - tests must pass, handoffs must validate - that you extend per project.
The examples directory has end-to-end recipes: a single-file refactor loop, a fan-out migration sweep, and a scoped pipeline. Start from the one closest to your problem and trim from there. The rule we live by: pick the simplest topology that ships, and climb to a more complex one only when the simpler one provably cannot handle the work.
Contributing
We are keeping the core small on purpose, but we welcome examples, recipes, and bug fixes. Read the contributing guide before opening a pull request, and please file an issue first for anything larger than a fix so we can agree on scope.
Building production agent systems? Talk to TunerLabs - we engineer agentic systems for businesses worldwide, and this starter kit is the same foundation we ship from. Star the repo, open an issue, or reach out and we will help you adopt it safely.
GitHub Repositories
Topics:
Master Claude Code
The complete architecture guide — Skills, Agents, Memory & the full Tools reference. Everything in one beautiful page.