Workflow guardrails for when AI coding agents go off-script

What it does

Babysitter is a JavaScript SDK that wraps AI coding harnesses like Claude Code, Codex, and Cursor into deterministic orchestration runs. You write your workflow as an async JavaScript function—defining tasks, quality gates, and human breakpoints—and the SDK ensures the agent executes exactly that sequence, recording every decision in an immutable event journal. It ships with an internal headless harness for CI/CD pipelines, but can also delegate subagent tasks back out to whichever AI CLI is installed on the system.

The interesting bit

Instead of treating the AI as the planner, Babysitter treats it as an untrusted worker that can only invoke operations the JavaScript process explicitly permits. The plugin system is equally unusual: plugins are markdown instructions or JS process files that the AI reads and executes, meaning a plugin can scaffold Terraform, install npm packages, or reconfigure your linter by driving the agent through the same guarded runtime.

Key highlights

• Multi-harness support: Claude Code (recommended), Codex CLI, Cursor, Gemini CLI, GitHub Copilot, Pi, and OpenCode, plus a built-in internal harness for headless automation
• Workflow authority is real code, not YAML config—quality gates are JavaScript conditionals and breakpoints are enforced pauses, not suggestions
• Event-sourced immutable journal records every decision for later retrospection and debugging
• Plugin marketplace covers security audits, testing gates, deployment pipelines, and CI/CD scaffolding
• Internal harness can discover and delegate subagent tasks to other installed AI CLIs at runtime, enabling multi-agent workflows from a single entry point

Caveats

• Most non-Claude harnesses (Cursor, Gemini, Copilot, Pi, OpenCode) are marked experimental; Codex CLI support is still in beta
• Requires Node.js 20+ and assumes you are already using an AI coding agent workflow; the internal harness works headlessly but the ecosystem is clearly optimized for Claude Code

Verdict

Worth evaluating if you are already using Claude Code or similar agents and need audit trails, mandatory human sign-offs, or repeatable CI/CD pipelines. If you are not yet using an AI coding harness, this is middleware for a stack you do not have.