Babysitter is an open-source orchestration layer on GitHub from a5c.ai that enforces discipline on AI agent workforces, letting coding agents manage extremely complex tasks through deterministic, hallucination-free self-orchestration. The project has about 1,900 stars and is distributed as an npm package with integrations for popular agent tools.
Autonomous coding agents are powerful but prone to drifting, skipping steps and declaring victory early. Babysitter wraps them in a strict process framework: work is broken into defined, verifiable steps, progress is tracked deterministically, and agents must satisfy each stage’s requirements before moving on. It ships as skills, workflows and plugins for environments such as Claude Code and Codex, so existing agentic setups gain structure without being replaced. The result is a way to run long, multi-part engineering workflows, sometimes described as babysitting your agents, where reliability comes from the process rather than from hoping the model behaves.
The project’s playful name captures a real frustration: even the best agents need supervision on long jobs. By encoding the supervision as reusable workflows instead of ad-hoc prompting, teams can share reliable agent processes the same way they share code.
Babysitter is free and open source under the MIT license. Install it from npm or GitHub and use it with your existing agent subscriptions and API keys; the orchestration framework itself has no paid tier.
Babysitter is for developers and teams who run AI coding agents on serious, multi-step projects and need those agents to follow a process reliably, from solo builders automating feature work to teams coordinating several agents at once.