Babysitter
Code and enforce agent-workforce workflows with approval gates, quality checks, and replayable run records.
- Source repo
- a5c-ai/babysitter
- Stars
- ★ 1.9k
- Last updated
- 2mo ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 59/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude CodeCodex · Claude.ai (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An engineering lead who needs a delivery sequence such as plan, human approval, implementation, testing, and score-based refinement encoded and enforced in JavaScript.
- Not a fit if
- Teams committed to a single coding tool and unwilling to install harness plugins
- Teams that reject mandatory breakpoints and human approval gates
- Source review
- 59/100 · Major gaps
What does this agent do, and when should you use it?
Babysitter is a workflow-orchestration project for AI coding harnesses, exposed through the babysitter, genty, and adapters CLIs plus harness plugins. Its workflows are JavaScript processes that can call ctx.task and ctx.breakpoint, with the runtime using that process to determine which next task is permitted. Tasks, gates, and decisions are recorded in an event-sourced journal, enabling replay and interrupted-run resumption. Version 6 uses the Adapters runtime to run the same process across supported coding harnesses, while an internal harness can execute processes for scripts, tests, and CI/CD without an external coding agent. The repository also includes blueprints, a process library, configurable compression, MCP server startup, and genty runtime-management commands.
A user starts a run through in-session commands such as /babysitter:call, /babysitter:yolo, and /babysitter:plan, or with genty call --harness <harness> --prompt <prompt> --workspace <path>. In a JavaScript process, await ctx.task(...) dispatches work and await ctx.breakpoint(...) creates a required human gate; normal code conditions can use quality-check results to gate subsequent work. After each step, the runtime checks what the process permits next and records tasks, gates, and decisions under ~/.a5c/runs/ by default, with compatibility reads from repository-local .a5c/runs/. The genty CLI provides resume, doctor, retrospect, cleanup, and start-server; the internal harness can execute a process or prompt directly and delegate subtasks to discovered installed harnesses. Plugins automatically register four compression layers, configurable through BABYSITTER_COMPRESSION_ENABLED, babysitter compression:toggle, and .a5c/compression.config.json.
- An engineering lead who needs a delivery sequence such as plan, human approval, implementation, testing, and score-based refinement encoded and enforced in JavaScript.
- A CI/CD maintainer who wants to run a lint-and-test process through genty call --harness internal where no external coding-agent CLI is available.
- A Claude Code team that needs structured approval gates during a long-running task and wants to resume it with /babysitter:resume after interruption.
- An engineering group using Claude Code, Codex CLI, or other supported harnesses that wants to reuse one process through Adapters.
- A team that must audit and review automated runs using journal records together with genty doctor and genty retrospect.
How do you install or deploy this agent?
Prerequisites are Node.js 20.0.0+ (the adapters CLI requires 20.9.0+) and a supported AI coding harness. Install the main CLI:
npm install -g @a5c-ai/babysitterTo drive a harness directly from the shell, install:
npm install -g @a5c-ai/adapters-cli
adapters doctorFor runtime orchestration commands, install:
npm install -g @a5c-ai/genty-platformFor Claude Code, add the a5c-ai/babysitter marketplace and install [email protected]. For Codex CLI (Beta), run codex plugin marketplace add a5c-ai/babysitter. The supplied material does not document an additional credential-configuration step.
How do you use this agent?
In a Claude Code session, run:
/babysitter:user-install
/babysitter:project-install
/babysitter:doctor
/babysitter:call implement user authentication with TDDA minimal runtime invocation is:
genty call --harness claude-code --prompt "implement user authentication with TDD" --workspace .
For automation without an external coding agent:
genty call --harness internal --process .a5c/processes/lint-and-test.js#process --workspace . --no-interactive --jsonResume an interrupted run with genty resume --run-id <runId> --harness claude-code --workspace .
What are this agent's strengths and limitations?
- Workflows are executable JavaScript rather than prompt-only conventions: ctx.task, ctx.breakpoint, and code conditions can constrain subsequent work.
- It records tasks, gates, and decisions in a journal designed for deterministic replay and resumption.
- It offers in-session plugins, the host-side adapters CLI, the genty runtime CLI, and an internal harness for both interactive and headless use.
- Structured human approvals, parallel dispatch, and quality gates provide governance mechanisms beyond a one-off task prompt.
- The main CLI requires Node.js 20.0.0+, while the adapters CLI has a higher 20.9.0+ floor.
- Except for the internal harness, typical use depends on an installed supported AI coding harness, with installation differing by harness.
- Codex CLI support is marked Beta; Cursor, Gemini CLI, GitHub Copilot, Pi, Hermes, Oh-My-Pi, OpenCode, and OpenClaw support are marked Experimental.
- Adopting the full stack requires understanding separate package roles for @a5c-ai/babysitter, @a5c-ai/adapters-cli, @a5c-ai/genty-platform, and harness plugins.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Babysitter This agent | 59 · Major gaps | CLIFree + model costs | ★ 1.9k | 2mo ago | JavaScript | Claude Code |
| Agency Orchestrator: One Sentence to Orchestrate an AI Expert Team | 71 · Some gaps | CLIFree + model costs | ★ 2.5k | 5d ago | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| Dagu Workflow Orchestrator | 65 · Some gaps | CLIFreemium | ★ 4.6k | today | Go | Codex · Claude Code · Claude API |
| DuraGraph | 50 · Major gaps | CLIFree | ★ 161 | 1mo ago | Go | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The security architecture explicitly disclaims default strong isolation and capability safety, and identifies filesystem, plugin, hook, network, and supply-chain risks. CI workflows grant contents, pull-request, and issue write permissions; conflict repair can force-push, so least privilege and external-effects controls are only partial. Interactive mode has approval breakpoints, while YOLO explicitly removes them. Journaling and recovery are described, but sensitive-data handling and dependency-risk mitigations are not. Maintainer names and a contact email appear in the repository, but publisher identity is unverified and cannot establish attribution.
Event journals, resume commands, diagnostics, and troubleshooting support ordinary use. However, the README claims deterministic, hallucination-free behavior while the security architecture limits stronger guarantees to specific implementations and validation; the installation material also disagrees about the minimum Node.js version. The product depends on external harnesses, npm packages, and remote services, whose availability is not established by the supplied material.
The materials cover interactive, autonomous, planning, continuous, and internal execution, with examples for multiple harnesses, CI, and scripts. They also state boundaries around external harnesses and isolation. Skill triggers are fairly specific, but harness maturity varies and some integrations are marked experimental; Node.js and dependency requirements constrain environment fit.
The documentation has a clear map of packages, install paths, commands, examples, and troubleshooting; the security architecture distinguishes current controls from deferred goals. The MIT license text is complete. Installation materials disagree on the minimum Node.js version, and the marketplace manifest says 6.0.2 while the root package.json says 6.0.3. The supplied sample has no verifiable changelog. Contact information is present, but maintenance responsibility and publisher identity remain unclear.
Process code, tasks, quality gates, breakpoints, and run journals provide a reviewable structure for complex workflows, with multi-harness orchestration as a clear source of added value. Commands and examples make outputs actionable, but the multi-package installation chain, external runtime dependencies, and integrations at different maturity levels add adoption and maintenance costs.
The documentation ties some architecture claims to named packages, commands, control surfaces, and validation requirements, and distinguishes current controls from deferred goals. The supplied materials do not directly substantiate the README's deterministic and hallucination-free claims, while the security document qualifies those assurances. Cross-file comparison reveals inconsistent version and environment requirements. Some capability and performance descriptions remain product claims that cannot be independently established from this static sample.
- YOLO mode has no approval breakpoints; ensure the task authorization and runtime environment are suitable for autonomous execution.
- The security architecture says strong isolation and capability safety are not provided by default; treat plugins, hooks, external harnesses, and remote services as trust boundaries.
- The installation materials disagree on the minimum Node.js version, and the supplied plugin manifest and root package show different versions. Check the target package's release notes before installing.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/babysitter)<a href="https://followagents.com/en/agents/babysitter"><img src="https://followagents.com/badge/babysitter.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
FAQ
Do I have to use Claude Code?
Can a workflow continue before a human approves it?
Is run state lost after an interruption?
Are pricing or additional credentials documented?
Thanks — we use this to improve the page.