Run a production deployment through a Codex orchestrator and subagents

A Redditor let Sol 6.1 High in Codex orchestrate a production deploy of a small app; it took almost 16 hours and over 90% of a ChatGPT Business allowance.

Done withCodex· Nov 12, 2026

What they did

A Redditor uses the Sol 6.1 High model in Codex as the main orchestrator for their software projects, coordinating specialized subagents for implementation, testing and auditing. They let it run a production deployment of a small application whose frontend and backend each build in about 25 seconds and whose PostgreSQL backup takes about 80 seconds. Afterwards they asked the agent for a technical audit of the deployment architecture, code, logs and history to find out why it took so long.

How it went

The deployment took almost 16 hours, after a previous one had taken about 14, and used more than 90% of the poster's ChatGPT Business usage allowance. The audit put over 95% of the time down to blockers, redundant checks and recovery steps the agent had added itself. The deploy framework had grown to more than 50 modules, with a 9,800-line main Python script, custom SQL parsers and repeated hash and schema checks; five of six release changes were fixes to the deploy tooling. In the comments the poster said a different model cut the run to about 20 minutes for around $40.

Worth knowing

Each time a deploy step failed, the agent added another safeguard or abstraction instead of fixing the cause, until the safeguards became the failures. Commenters asked why it was left running for 16 hours, and suggested committing runbooks and deploy history so the agent can see how long a deploy should take. One commenter reported the same behaviour. Check in early and cap how long an agent-run deploy may take.

Try it

With Codex· 5 parts to fill in

Deploy [my app] to [production environment] using the existing pipeline in [repo or runbook]. Past deploys take about [normal duration]; if this one runs past [time limit], stop and tell me what is blocking it. Fix only the immediate cause of a failure, do not add new validation layers, modules or permanent safeguards, and do not rerun stages that already passed. Before changing any deployment code, show me the diff and wait for my approval.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related