a5c.ai
Babysitter

Stop babysitting your agents. Start shipping.

The only way to ship complex tasks with AI agents.
Complex work that actually gets merged.
Tasks converge
PRs get merged
"Almost done" becomes done
100% open source
Install Babysitter on Claude Code
# Add the plugin repository
claude plugin marketplace add a5c-ai/babysitter
# Install the plugin
claude plugin install --scope user babysitter@a5c.ai
Then run /babysit + your task
Star us on GitHubWhy this works →
Alpha·Break things with us
a5c babysitter
● Agent(planning)
  ⎿ Executed for 7 iterations and 10h 48m 35s. Score: 82/100.

● Orchestration Step (Move on to next phase)
  ⎿ Moved on to next phase.

● Agent(building tests)
  ⎿ Executed for 3 iterations and 2h 30m 20s. Score: 95/100 (Passed).

● Orchestration Step (Breakpoint approval)
  ⎿ Breakpoint approval requested and released.

● Agent(developing code)
  ⎿ Executing for 5 iterations and 4h 30m 20s. Score: 75/100 (4/5 tests passed)
    ⎿ Fixing remaining gaps and test failures.

✽ Babysitting… (3d 20h 4m 46s · ↑ 3.6m tokens)

> /babysit "use a test driven process to build a backoffice for this repo"

You already use AI agents.

You’re just not satisfied with them.

You’ve tried:
Cursor
Copilot / agent modes
CLI agents (Claude, Codex, Gemini)
“Autonomous” dev tools
They’re impressive.
And yet:
they break on complex tasks
they lose context
they stop when things get messy
they leave you to retry, review, and stitch
they don't verify work meets your standards
You don’t fully trust your agents.
Even simple refactors need constant hand-holding.
Install nowStar us on GitHub

Get much better results with YOUR agents.

Complex tasks that break other agents? Babysitter handles them.

You keep:
Claude code
Codex
Gemini
whatever coding agent you trust
Babysitter wraps them in a system that:
is Git-native
retries intelligently
enforces tests and quality gates
handles failure paths
knows when to stop
knows when to ask you
works on tasks from simple to complex
Same agents.
Radically better outcomes.
Install nowStar us on GitHub

Babysitter runs the process. Your agent does the work.

Agents alone are unpredictable. Babysitter keeps them on track.

Babysitter

Runs the process
  • Plans the steps
  • Checks the results
  • Retries until it passes
“Keeps it on track”

Your Agent

Claude, GPT, any LLM
  • Writes the code
  • Runs the commands
  • Fixes the errors
“Does the actual work”
Install nowStar us on GitHub

Why Babysitter gets better results than dev agents alone

Most dev agents:
optimize for responsiveness
stop on first failure
treat complexity as an edge case
Babysitter:
optimizes for quality and completion
expects failure and recovers
is built for long, multi-step, messy tasks
That’s why:
tasks converge instead of drift
PRs get merged instead of abandoned
“almost done” becomes “done”
This isn’t smarter prompting.
It’s owning control over the execution.
Install nowStar us on GitHub

What is Babysitter

Think of it as a the manager for your AI agents. It keeps them on track so you don't have to.

You configure it as an agent skill in Claude Code.

  1. 1.
    You define:
    Human
    Your expected outcome (e.g., "add dark mode" or "refactor auth")
  2. 2.
    Babysitter creates processes and methodology
    Babysitter
    Defining the process as structured deterministic codequality gates (e.g. tests)human breakpoints (predetermined points for human intervention)
  3. 3.
    You approve or change/update the process
    Human
  4. 4.
    The Babysitter begins orchestrating:
    Babysitter
    streamlines your agentretries on failurerecovers automaticallypauses only when neededstops only when “done” is satisfied
  5. 5.
    You can finally take that lunch break that you missed
    Human
Install nowStar us on GitHub

This is for you if:

You already believe in AI for development
You’re tired of babysitting agents
You want to get more out of ai engineering
You are a Ralph Wiggum loops fan
You work on tasks that don’t fit in one prompt
You want speed without losing control
You stopped working on very complex tasks with agents
If Cursor and Claude code feels lazy
and your CLI agent feels fragile -
The babysitter is the missing layer.
Install nowStar us on GitHub

Bring your own agent.

Get real results.
Ship complex work.
Most agent tasks fail at 80% done.
Babysitter gets you to 100%.
Concrete behavior
Runs your agent inside your workflow + quality gates
Writes an artifact trail into .a5c/runs/
Stops at breakpoints instead of guessing
User GuideSee Agent Stories →
Ralph Loops

Ralph loops are an amazing hack

They automate repetition, not execution ownership.

The Babysitter:execution ownership
enforces quality gates, tests, and checkpoints
retries intelligently only on real execution signals
pauses for human judgment where it matters
stops when work is verifiably done
Key contrast:
Ralph loops keep your agent working.
Babysitter enforces YOUR quality standards.
Install nowStar us on GitHub
Hall of Fame

Complexity Legends

PRs that shipped while you were sleeping

15K+ Lines

Codex Subagent Architecture

a5c-incubator/codex

0additions
0files
24 hoursunattended
New codex-subagent crate
Agent registry & runtime
Full test suite
View on GitHub
Full Platform

Hub - Git Hosting Platform

a5c-ai/hub

Full Platform
30+ PRs
GitHub-like functionality
Azure + Terraform
Actions runners & webhooks
View on GitHub

Your PR Could Be Here

Shipped something complex with Babysitter?

Get featured in the Hall of Fame

Submit Your PR

Can your agent do this?

15K+ lines. 159 files. Zero human intervention.

Try Babysitter
Explore More Agent Stories

Join the Community

Connect with builders creating processes together: sharing workflows, quality gates, and the hard-won details that make complex tasks succeed.

Star us on GitHubJoin Discord