← Back to research
·8 min read·company

Babysitter

Babysitter is an orchestration framework for AI coding agents — Claude Code first, with beta/experimental support for Codex, Cursor, Gemini CLI, and others — enabling deterministic, event-sourced workflow management with quality gates, human-in-the-loop approval, and resumable runs.

Key takeaways

  • Deterministic, event-sourced orchestration framework for AI coding agents with resumable runs and complete audit trails
  • Quality convergence — agents iterate until quality targets are met, not just run once and hope
  • No longer Claude Code-only — Codex CLI support is in beta, with experimental support for Cursor, Gemini CLI, Copilot, and OpenCode
  • Traction quadrupled in 2026 — 1,310 GitHub stars and ~9,100 weekly npm SDK downloads as of June 2026, though no new release since April 4

FAQ

What is Babysitter?

Babysitter is an orchestration framework for AI coding agents — Claude Code first, with beta Codex CLI support — that enables deterministic, event-sourced workflow management with quality gates, human approval checkpoints, and automatic iteration until quality targets are met.

How does Babysitter work?

Babysitter runs an iterate-execute-record loop: advance process, get pending effects, execute tasks, post results, repeat until complete. Everything is event-sourced in .a5c/runs/ for full resumability and audit trails.

How much does Babysitter cost?

Babysitter is free and open-source under the MIT license, with zero telemetry. An enterprise offering is advertised on a5c.ai but no pricing is disclosed. You pay for the underlying agent (Claude Code or another supported harness).

Who competes with Babysitter?

Claude-Flow for swarm orchestration, wshobson/agents for plugin-based multi-agent coordination, and BMAD Method for full agile lifecycle methodology.

Executive Summary

Babysitter is an orchestration framework that brings deterministic, event-sourced workflow management to AI coding agents — Claude Code first and foremost, with beta support for Codex CLI and experimental support for Cursor, Gemini CLI, GitHub Copilot, and OpenCode. Unlike frameworks that provide static skills or methodology guidance, Babysitter actively manages multi-step workflows with quality gates, human approval breakpoints, and automatic iteration until quality targets are met. A5C now positions it as "CI/CD for AI agents." [1] [2]

AttributeValue
CompanyA5C AI
Founded2026
FundingNot disclosed
LicenseMIT
GitHub Stars1,310 (as of June 11, 2026)

Product Overview

Babysitter sits between methodology frameworks (which tell agents what to do) and runtime orchestrators (which manage how agents execute). It provides an event-sourced execution loop where every step is recorded, resumable, and auditable. [1]

The core innovation is quality convergence: instead of running a workflow once, Babysitter iterates until quality targets are met, with agent scoring and parallel execution support. Human-in-the-loop breakpoints provide structured approval gates — not just "approve/reject" but context-rich checkpoints.

Key Capabilities

CapabilityDescription
Event-Sourced ExecutionComplete journal of all events in .a5c/runs/
Quality ConvergenceIterate until quality targets met, not one-shot
Human BreakpointsStructured approval gates with full context
Parallel ExecutionRun tasks concurrently with dependency management
Deterministic ReplayResume or replay any run from any point
Process Library2,000+ pre-built process definitions
Agent ScoringEvaluate agent output quality programmatically
Token Compression4-layer compression claiming 50–67% token reduction [1]

Installation

On Claude Code, Babysitter installs as a plugin; other harnesses install via npm or native installers: [1] [3]

npm install -g @a5c-ai/babysitter@latest @a5c-ai/babysitter-sdk@latest @a5c-ai/babysitter-breakpoints@latest
claude plugin marketplace add a5c-ai/babysitter
claude plugin install --scope user babysitter@a5c.ai
claude plugin enable --scope user babysitter@a5c.ai

Invoke via /babysitter:call or natural language in Claude Code.


Technical Architecture

Babysitter runs an iterate-execute-record loop:

  1. Advance process — Determine next steps from process definition
  2. Get pending effects — Identify tasks that need execution
  3. Execute tasks — Run tasks (potentially in parallel) via Claude Code subagents
  4. Post results — Record outcomes to event journal
  5. Repeat — Continue until all tasks complete or quality gates pass

Key Technical Details

AspectDetail
RuntimeNode.js 20+ (22.x LTS recommended)
AgentClaude Code (recommended); Codex CLI (beta); Cursor, Gemini CLI, GitHub Copilot, OpenCode (experimental); internal harness needs no external agent [1]
StateEvent-sourced journal in .a5c/runs/<runId>/
ExecutionSequential and parallel task execution
LanguageJavaScript
Open SourceYes (MIT)

Strengths

  • Deterministic workflows — Event-sourced execution means runs are reproducible and resumable
  • Quality convergence — Iterates until quality targets are met, unlike one-shot frameworks
  • Human-in-the-loop — Structured breakpoints with rich context for meaningful approval gates
  • Audit trail — Complete journal of every event, task, and decision
  • Large process library — 2,000+ pre-built processes cover common development workflows
  • Composable — Works with existing Claude Code subagents, skills, and tools
  • Broadening runtime support — No longer Claude Code-only: Codex CLI in beta, four more harnesses experimental [1]
  • Active development — Repository pushed June 11, 2026; growing fast (1,310 stars, 75 forks) [1]

Cautions

  • Claude Code-first in practice — Multi-harness support exists, but only Codex CLI is beyond experimental; everything else is labeled experimental [1]
  • Release cadence stalled — Latest release (v0.0.187) and npm publish date to April 4, 2026, two months before this update, even though repository commits continue [1] [3]
  • Still early — Five months old; limited public production validation despite star growth
  • Complexity — Event-sourced architecture adds overhead vs. simpler skill-based approaches
  • Process library quality — 2,000+ processes remains a bold, largely unverified claim
  • Unknown company — A5C AI still has no disclosed funding, team, or pricing; an enterprise page exists but with no details [2]

What Developers Say

No substantive attributed developer commentary surfaced in searches as of June 11, 2026 — no Hacker News launch thread, no notable Reddit discussion. Coverage is limited to aggregator listings and A5C's own materials. For a project that quadrupled its stars in four months, the absence of independent practitioner write-ups is the notable signal: people are starring it, but few are publicly reporting production experience.


Pricing & Licensing

TierPriceIncludes
Open SourceFreeFull framework, MIT license, zero telemetry [2]
EnterpriseNot disclosedAdvertised on a5c.ai; no public details [2]

Hidden costs: Requires a Claude Code subscription or Anthropic API costs (or the equivalent for another supported harness)


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
Claude-FlowClaude-Flow does swarm orchestration with consensus protocols; Babysitter does deterministic event-sourced workflows with quality convergence
wshobson/agentswshobson/agents is plugin-based multi-agent coordination; Babysitter is process-driven single-agent orchestration
BMAD MethodBMAD provides agile lifecycle methodology; Babysitter provides runtime workflow execution with quality gates
SuperpowersSuperpowers enforces methodology via skills; Babysitter enforces via deterministic process execution

When to Choose Babysitter Over Alternatives

  • Choose Babysitter when: You need deterministic, resumable workflows with quality convergence and audit trails in Claude Code (or, with beta support, Codex CLI)
  • Choose Claude-Flow when: You need distributed multi-agent coordination with formal consensus
  • Choose wshobson/agents when: You want composable plugin-based agents with broad tool coverage
  • Choose Superpowers when: You want methodology enforcement without runtime orchestration overhead

Ideal Customer Profile

Best fit:

  • Teams using Claude Code who need structured, multi-step development workflows
  • Organizations requiring audit trails and approval gates for AI-generated code
  • Developers working on complex features that benefit from iterative quality improvement
  • Teams that want "iterate until right" instead of "run once and fix manually"

Poor fit:

  • Teams standardized on harnesses where support is still experimental (Cursor, Gemini CLI, Copilot, OpenCode)
  • Developers wanting lightweight skills without orchestration overhead
  • Organizations that need a vendor with disclosed funding, team, and pricing
  • Quick one-off tasks where orchestration is overkill

Viability Assessment

FactorAssessment
Financial HealthUnknown — no disclosed funding
Market PositionNiche — Claude Code-first orchestration, expanding to other harnesses
Innovation PaceActive commits as of June 11, 2026, but no release since April 4 [1] [3]
Community/EcosystemGrowing — 1,310 stars, 75 forks, ~9,100 weekly SDK downloads on npm (as of June 2026) [1] [3]
Long-term OutlookUncertain — improving traction, but A5C AI remains opaque

The event-sourced, quality-convergence approach is genuinely novel in the skills framework space. Most competitors either provide static methodology (Superpowers, BMAD) or distributed coordination (Claude-Flow, wshobson). Babysitter's niche — deterministic workflow execution with iterative quality — is underserved, and the runtime lock-in risk has eased now that Codex CLI support is in beta. The remaining risk is the unknown viability of A5C AI itself.


Bottom Line

Babysitter brings software engineering rigor — event sourcing, deterministic replay, quality convergence — to AI agent orchestration. The approach is more sophisticated than static skills frameworks and more focused than distributed multi-agent platforms.

Recommended for: Claude Code (and increasingly Codex CLI) users who need structured, auditable, iterative workflows for complex development tasks.

Not recommended for: Teams wanting lightweight skill catalogs, mature multi-harness support, or a vendor with a public track record.

Outlook: Traction quadrupled in four months (317 to 1,310 stars) and runtime support is broadening beyond Claude Code, both good signs. But the two-month release gap, absent public user commentary, and A5C AI's continued opacity mean the quality-convergence model still lacks independent production validation. If that validation arrives, expect the pattern — iterate until quality targets are met — to be adopted by larger orchestration platforms.


Research by Ry Walker Research • methodology