# Has anyone automated a ChatGPT \<\> Codex orchestration loop?

**URL:** <https://community.openai.com/t/has-anyone-automated-a-chatgpt-codex-orchestration-loop/1397120>\
**Category:** Use cases and examples\
**Tags:** codex\
**Created:** [September 13, 2026, 9:39am UTC](https://community.openai.com/t/has-anyone-automated-a-chatgpt-codex-orchestration-loop/1397120 "2026-09-13T09:39:52Z")\
**Posts on this page:** 2\
**Page:** 2

<div class="post-metadata">

**Author:** ![Pimpcat](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/pimpcat/32/738039_2.png) [@Pimpcat](https://community.openai.com/u/Pimpcat)\
**Post date:** [September 30, 2026, 1:05pm UTC](https://community.openai.com/t/has-anyone-automated-a-chatgpt-codex-orchestration-loop/1397120/23 "2026-09-30T13:05:41Z")

</div>

The “planner → agent → reviewer” summary leaves out most of the supervisor responsibilities in my original post.

Before launching a worker, the supervisor selects a bounded gate and creates a work order with scope, constraints, budget and acceptance criteria. Afterwards, it independently verifies the evidence and decides whether to advance, retry or hold. Durable state lives outside the worker conversation.

Choosing between human, deterministic and agent execution, and assessing worker qualification, are additional questions worth discussing. But work allocation, constraints, independent verification and state reconciliation were already part of the loop I described.

Here’s the workflow running on an active build. It’s catching failures, repairing them and checking acceptance before promotion. This packet is still in progress.

 ![image](https://us1.discourse-cdn.com/openai1/original/4X/4/f/8/4f82cfae8bfaa77404427cf7b951ecaf03c04eab.png)

---

<div class="post-metadata">

**Author:** ![miguelsanchez](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/miguelsanchez/32/767575_2.png) [@miguelsanchez](https://community.openai.com/u/miguelsanchez)\
**Post date:** [October 6, 2026, 10:40am UTC](https://community.openai.com/t/has-anyone-automated-a-chatgpt-codex-orchestration-loop/1397120/24 "2026-10-06T10:40:21Z")

</div>

the thing that matters most is making the repo the only source of truth. keep a plan file plus one small packet file per task in the repo, with the goal, the files it may touch and an acceptance command. the planner reads git state and those files, never the chat history. each Codex worker gets one packet, runs it non-interactively in a fresh checkout and writes a receipt back with the diff, the commands it ran and their exit codes. then the supervisor runs the acceptance command again itself instead of trusting the worker’s summary. most of the silent drift comes from a worker reporting tests passed when they never ran. make the human gate a status field in the packet that the loop refuses to move past, rather than something the model decides. cap retries per packet, and give each retry a new worker with only the previous failure log attached so it doesn’t inherit a confused context. GitHub Actions is fine as the trigger and audit trail. start with a dumb script loop before reaching for an agents framework, it’s much easier to debug

[Previous page](https://community.openai.com/t/has-anyone-automated-a-chatgpt-codex-orchestration-loop/1397120.md?page=1)