Engineers at OpenAI showed Codex agents that run for days and weeks from a short goal instead of a long plan.
Can you actually leave a coding agent alone that long? Friel said yes, with the right outcome and checks, after a CI agent ran for more than 1,000 hours.
Vivi started with personal builds to prove the point. She gave Codex with Astra seven reference photos and got back an explorable Ferrari model with clay and wireframe versions, then tasked it with only a track name and got a 21 kilometer Nurburgring fly-through with all 70 turns built over two days, mostly waiting on Blender renders. Friel pushed the same idea into production work where continuous integration is the automated tests that decide if code can merge, and a flaky test is one that fails once then passes without a code change. His 200 word prompt told the agent to watch CI, reproduce failures, propose fixes and open pull requests, and over six weeks it opened more than 100 pull requests per week, pushed more than 1,000 commits, sent more than 600 Slack coordination messages and merged 213 fixes, Friel said.
So how did it not stall or wander? About four days in Friel told it to stop doing everything in one thread and let subagents grind through hard failures while the main thread triaged, launched tests and messaged owners, a change that unlocked 5,800 subagent sessions and just under half a million tool calls with more than 67,000 waits for builds and test results, Friel said. The hardest win was a non deterministic floating point error in a GPU test that took days to reproduce on one node, then required direct questions to the owning researchers about software versus hardware fault. Vivi said the scaffolding that makes that durable is one thread per outcome with a goal it chases, an agents.md file with dos and hard limits plus a running work log, built in compaction so three month threads do not lose context, remote machines or Codex Cloud with snapshotted environments, plugins for GitHub and Slack context, child threads with separate worktrees so parallel edits do not collide, plus ultra mode for one hard bug and heartbeat checks that revisit a pull request every half hour.
That still leaves cost.
Friel said the six week run cost about $32 per hour in tokens, now about $16 per hour on GPT-6 Sol, which stretches the token budget twice as far, with 6.1 Sol more efficient still, according to his pricing math shared at OpenAI launches GPT-5.6. Astra, the model family behind that work, also powers GPT-6 Pro and runs inside ChatGPT Work and Codex where it can operate the same applications people use every day, according to OpenAI. For teams that own CI health, the workflow change is concrete: write the outcome and merge rules once, give the agent read access to tests and issues, budget tens of dollars per hour instead of engineer days, and watch whether on call time shifts from chasing flakes to reviewing agent pull requests.
