a5c-ai/babysitter
Babysitter enforces obedience on agentic workforces and enables them to manage extremely complex tasks and workflows through deterministic, hallucination-free self-orchestration
What it solves
Babysitter prevents AI agents from hallucinating or deviating from a defined workflow. It provides a deterministic orchestration layer that ensures agents follow a strict sequence of steps, pass quality gates, and stop for human approval at specific breakpoints, rather than relying on the agent's own unpredictable planning.
How it works
Workflows are defined as JavaScript code (processes), which act as the absolute authority for the agent's actions. The system uses an enforcement mechanism that mandates stops and checks against the process definition. It is harness-agnostic, meaning it can be integrated into various AI coding tools (harnesses) via adapters, or run headlessly using an internal harness for CI/CD pipelines.
Who it’s for
Developers and teams managing complex agentic workflows who need high reliability, strict adherence to process, and human-in-the-loop oversight for AI-driven coding tasks.
Highlights
- Harness-Agnostic: Supports 12 different AI coding harnesses, including Claude Code, Codex, Cursor, and GitHub Copilot.
- Deterministic Orchestration: Uses real code to define processes, ensuring agents do exactly what is permitted and nothing more.
- Human-in-the-Loop: Enforces mandatory breakpoints for human approval before proceeding.
- Immutable Journaling: Records every decision and action in a journal for traceability.
- Internal Harness: Includes a built-in engine for headless execution in scripts and CI/CD pipelines without requiring an external AI agent.