Set the goal.Leave the loop.
OpenAI Build Week 2026 project↗
Set the goal.Leave the loop.
Plan once in ChatGPT. AgentLoop sends the goal to fresh Codex workers, reviews every result with an independent critic, and improves your local project until it passes your standards.
Try the live replay ↗Watch the 3-minute demo
Fresh workersIndependent criticFiles as memory
127.0.0.1:5757
LIVE
CRITIC / 01FAILConcrete fixes written
CRITIC / 02PASS11 tests passing
01
Fresh context every cycle
02
Files carry the work forward
03
Your rubric decides done
01 / THE SYSTEM
Stop being the relay between planning and code.
Set the goal and standards once. AgentLoop handles every handoff: planner to worker, worker to critic, then critic feedback into the next cycle. You stay in control without carrying every message.
01PlanDefine the finish line02BuildStart with clean context03CritiqueGrade the work, independently04RepeatCarry evidence forward
01 / Plan
Define the finish line
Plan in ChatGPT, then send AgentLoop a bounded goal. Your rubric stays in GUIDELINES.md.
cycle.loglive
01goal.md written
02GUIDELINES.md loaded
03cycle budget: 3
02 / START LOCAL
Run it where your code lives.
Requires Node.js 18+, Git, and an authenticated Codex CLI. No npm install. No hosted workspace.
WindowsmacOSLinux
Copy
$`git clone --branch v1.1.0 --depth 1 https://github.com/aiedwardyi/AgentLoop.git`
›`cd AgentLoop`
›`node src\daemon.js`
Stable release v1.1.0MIT licensedLocal-first
03 / GIVE IT WORK
Use the loop for work you would otherwise babysit.
Give it a feature, migration, hardening pass, or product polish with a bounded goal and clear acceptance criteria. Fresh workers build, independent critics review, and project files carry the evidence forward.
01Feature02Migration03Hardening04Polish
Example goal
Ship a complete feature
`Add CSV export across the UI, API, and regression suite.`
Critic checks
01Behavior works end to end02Edge cases are covered03Docs match the resultPASS, then polish to SHIP
04 / THE RECEIPT
A critic catches what green tests miss.
In a separate query-parser evaluation with no forced failure, all nine tests passed. The fresh critic still found a real mixed percent-decoding defect.
Cycle 01FAIL
9 / 9 tests passed
Fresh critic found mixed percent-decoding that the existing suite missed.
`fix: preserve malformed escape sequences`
Cycle 02PASS
11 / 11 tests passed
The next worker fixed the defect, added regression coverage, and cleared the same rubric.
`VERDICT: PASS`
05 / OBSERVABILITY
Run it unattended without losing the plot.
Watch every cycle, verdict, transcript, cost, and file handoff. Cancel from the dashboard, or ask ChatGPT for status through MCP.
Open the interactive replay ↗Watch the full demo ↗
WorkerCriticControlREAL RUN
Fresh process / cycle 2
Work happens in the real project
A clean worker reads the goal, edits local files, runs tools, and leaves evidence behind.
06 / YOUR STANDARD
Prompts disappear. Standards stay.
Put your definition of done in GUIDELINES.md. Every worker builds against it. Every fresh critic enforces it.
GUIDELINES.mdREAD EACH CYCLE
01
All tests must pass.
02
Preserve malformed input without data loss.
03
Add regression coverage for every defect.
final lineVERDICT: PASS
OPEN BY DESIGN
Codex ships first. The engine boundary stays pluggable.
MCP keeps planning separate from local execution. Next: more engines, two-way agent questions, and research loops.
THE RELAY IS GONE