Open sourceBabysitter
Babysitter. Your workflow, actually enforced.
Define your workflow in code. Every step is enforced, quality gates pass before progression, humans approve at breakpoints, every decision is journaled.
01Install
One install, then your harness.
The harness (agent, skills, tools, process, memory) is built around a coding agent: Claude Code, Codex, Cursor, and 9 others. Every harness shares the same first step.
npm install -g @a5c-ai/babysitterclaude plugin marketplace add a5c-ai/babysitter-claude
claude plugin install --scope user babysitter@a5c.aicodex plugin marketplace add a5c-ai/babysitter-codex
codex plugin add babysitter --marketplace babysitterMore harnesses
The other ten harnesses install through the universal helper. The argument is the harness key, which is not always the harness name.
babysitter harness:install-plugin <harness-key> [--workspace <path>]- Cursor adds the marketplace through its UI; there is no CLI marketplace add.
- Antigravity and Gemini install via the SDK helper only; there is no standalone npm installer.
- genty installs from npm only.
Requires Node.js 20+; 22.x LTS recommended. The README labels the Codex integration “(Beta)”. Commands verified July 2026 against the harness install matrix.
One install, and your next run keeps a journal you can replay.
02Mechanism
How it works.
Two loops. The deterministic loop holds state, budgets, gates, and the journal; the adaptive loop reasons inside it. Your agents do exactly what the process permits, nothing more. The README puts it bluntly: enforce obedience on agentic workforces.
Process as code
The workflow is JavaScript. The orchestrator can only do what this code permits: ctx.task, ctx.breakpoint, gate logic written as real code.
Mandatory stop
The harness cannot keep running on its own. Every step ends with a forced stop; the process decides what comes next.
Enforcement
Gates block progression until they are satisfied. A step that fails its gate does not proceed.
Event-sourced journal
All run state lives in ~/.a5c/runs/ as an append-only record, with deterministic replay and resume from any point.
03In code
The shape of a process.
A process is JavaScript. Tasks run through ctx.task, humans approve at ctx.breakpoint, and a gate is an ordinary conditional the run cannot argue with.
// Illustrative process definition. Gate logic is plain code.
export default defineTask("draft-and-review", async (ctx) => {
let draft = await ctx.task("write-draft");
// Quality gate: the run cannot progress until the gate passes.
let review = await ctx.task("review-draft", { draft });
while (!review.pass) {
draft = await ctx.task("revise-draft", { draft, feedback: review.feedback });
review = await ctx.task("review-draft", { draft });
}
// Breakpoint: the run stops here until a human approves.
await ctx.breakpoint("editor-approval", { draft, review });
return draft;
});An illustration, condensed. The full API and real process definitions are in the repository.
04The contract
What you get.
- 01Process enforcement
- Steps run in the order the code allows, and no other order.
- 02Deterministic replay
- Every run reconstructs from its journal and replays to the same decisions.
- 03Breakpoints
- The run stops and waits for a human before consequential steps.
- 04Parallel execution
- Tasks fan out under the same process; the journal keeps one record.
- 05Quality convergence
- Multi-dimensional quality scores, iterated until the target is met (the docs' 90-Score Pattern: a scoring method, not a benchmark result).
05Portability
Harness-agnostic as of v6.
The same process runs across the 12 supported harnesses, on the Adapters runtime. Write it once; the journal, gates, and breakpoints follow it to every harness.
06Modes
Five ways to run.
- 01call
- Interactive. Breakpoints stop the run for approval.
- 02yolo
- Autonomous. The same process, no breakpoints.
- 03plan
- Plan the run without executing it.
- 04forever
- Continuous runs that pick up new work.
- 05doctor
- Diagnostics for a run and its journal.
07Headless
The internal harness.
A built-in harness runs processes programmatically, with no external coding agent: the entry point for CI and headless operation. At runtime it can delegate tasks to any discovered installed harness, so a single headless entry point can drive a multi-agent workflow.
08Evidence
When is it worth it?
Not every task needs an enforced process. We scored all 163 Terminal-Bench tasks against a benefit-versus-cost rubric and published every judgment, including the ones where a plain agent is the better call.
163 tasks · Terminal-Bench 2.1 + 3
- 59Babysitter
- The enforced process earns its overhead.
- 61Borderline
- Either approach works; a judgment call.
- 43Vanilla
- A plain agent is the better call.
Longer tasks earn the overhead: on Terminal-Bench 3, 54% of tasks route to Babysitter, against 21% on the shorter 2.1, whose median task is a quarter the length.
Each task is scored on 13 dimensions, eight of benefit and five of cost, with a cited span of evidence behind every non-zero score and a counterfactual required before any benefit counts. Borderline cases are re-judged by a panel and settled by majority. The rubric and every judgment are open.
Babysitter is the enforcement engine of the a5c stack, one of its two pillars. Deterministic memory is the other: KIP, pre-release.
Want this operated for your domain? Request a demo.