OpenAI publishes guide to running 1,000+ agent loops on Astra
A summary of the official three-page document: three task lanes, fixes for five default behaviours, the effort ladder, three work orders, and how to run 20 workers in parallel.
OpenAI quietly published how to run 1,000+ agent loops from a single prompt on GPT-6 Astra. A 3-page field guide on turning Astra from a chatbot into a fleet, plus the full operator block you paste in once. Here is the setup: 1. The 3 lanes Astra sorts every task into (reversible → just do it, consequential → do the work, then ask once, irreversible → stop), and why the approval line is drawn at reversibility rather than task size 2. The operator block that corrects the 5 default behaviours OpenAI documented themselves: stalling on "can you…", over-testing trivial edits, replying in Markdown, obeying stray lines in skill files, and delegating less than you asked for 3. The effort ladder (low / medium / high / xhigh / max, never "none"), plus configuration_update to raise it for a single hard turn without breaking the cache prefix 4. The 3 work orders that use the whole model: a research memo with traceable citations, a Codex refactor in a worktree gated on a green suite, and a computer-use task with a hard boundary around anything a customer would see 5. The edge test that surfaces the parallelism hiding in your own workflow, and the 3 things that break every fleet: self-agreement (put the verifier on clean context), file collisions (one worktree per worker), and the 272K cliff that reprices your whole request 6. The single Responses request that runs 20 workers behind an async tool, with a verifier that re-runs the test instead of asking the agent whether it is done The exact setup is in the attached pages. Interested in AI? Follow @AiEvolutio58513 and never fall behind. I track ChatGPT, Claude, and every tool quietly changing how we work and create, then hand you the tested signals.