I Built an Open-Source Agent Kit for OpenClaw: Evaluation-Driven Development in Action

In this off-the-beaten-path episode, I step away from my BMAD experiments to walk you through a new open-source @GitHub repo: OpenClaw Agents, a compact but rigorous starter kit for building agents that work with Anthropic’s OpenClaw, not as a fork or clone of it.

I explain how conversations at work about evaluation-driven development (EDD)—a cousin of test-driven development—sparked the idea: if everyone’s talking about OpenClaw, why not build a well-structured agent layer around it, with tests and evals baked in from the start. Instead of chasing hype, I focus on a simple but disciplined repo that lets you define agents as markdown, run them through Python-based tests, and evaluate over time how well they actually perform.

You’ll see how I approach this as an orchestrator, not a traditional software engineer: designing an orchestrator agent that routes work to tightly focused subagents, each with narrow guardrails and clear responsibilities. I also show the extensive documentation and diagramming (in my private build repo) that shaped the agent architecture, and why I refuse to ship anything I wouldn’t personally use in my own workflows.

I walk through the repo structure—evals, tests, agent templates, and subagent templates—and talk about how I leaned on community standards to improve my open-source hygiene: code of conduct, contributing guidelines, readiness tests, TDD expectations, and lightweight CI validation. This project became a forcing function for me to level up as a repo maintainer so contributors have a clear, safe path to extend and customize the agents.

Throughout the video, I frame OpenClaw Agents as a complement to OpenClaw itself: a practical foundation for people who want to experiment with agent design, orchestration, and oversight while the ecosystem and Anthropic’s roadmap continue to evolve. I close by sharing why I believe the future of software development leans heavily on supervision, orchestration, and evaluation around LLMs—not replacing professional developers, but changing how we collaborate with these models—and how this repo is my contribution to that emerging way of working.

Compliments to the @Descript team as I used Underlord to help produce this video.

Видео I Built an Open-Source Agent Kit for OpenClaw: Evaluation-Driven Development in Action канала Tim Dickey