If your AI agent keeps going off the rails, Raindrop AI has a new open-source tool to help it course-correct on its own. Meet Workshop, a local debugger that integrates deeply with code agents like Claude Code and Cursor. It lets large language models directly read the agent's runtime traces, automatically write evaluation tests, and modify production code — forming a self-healing loop.
Workshop provides a local dashboard that streams every token, tool call, and decision in real time — no polling required. It also includes a local replay mechanism: developers can generate an HTTP endpoint from the command line to replay production traces directly in their local environment.
Here’s how it works in practice: when a business agent goes awry, developers can summon a code agent right from the terminal. With Workshop hooked in, the code agent reads the logs, writes test scripts for the failing scenario, and patches the code. Then it automatically retries, looping through evaluation and repair until all tests pass.
Workshop is now open source under the MIT license. It supports TypeScript, Python, Go, and Rust, and integrates with major frameworks like Vercel AI SDK, LangChain, and CrewAI.