A field guide covering operator configuration, task routing by reversibility, effort ladders, and parallel worker patterns for running 1,000+ agent loops from a single prompt.
Adapted from @adiix_officialOpenAI quietly showed how to run 1,000+ agent loops from one prompt on GPT-6 Astra 3-page field guide on exactly how to turn Astra from a chatbot into a fleet, along with the COMPLETE OPERATOR BLOCK you paste once here is how you set it up: 1. the 3 lanes of action Astra actually sorts every task into (reversible → just do it, consequential → do the work then ask once, irreversible → stop) and why the approval line is reversibility, not task size 2. the operator block that fixes the 5 default behaviours OpenAI documented themselves: stalling on "can you…", over-testing trivial edits, answering in Markdown, obeying stray lines in skill files, and delegating less than you want 3. the effort ladder (low / medium / high / xhigh / max, no "none") with configuration_update to raise it for one hard turn without breaking the cache prefix 4. the 3 work orders that use the whole model: research memo with traceable citations, Codex refactor in a worktree with a green-suite gate, computer-use task with a hard boundary on anything a customer would see 5. the edge test that finds the parallelism hiding in your own workflow, plus the 3 things that break every fleet: self-agreement (verifier on clean context), file collisions (one worktree per worker), the 272K cliff that reprices your whole request 6. the one Responses request that runs 20 workers behind an async tool with a verifier that re-runs the test instead of asking the agent if it's done this is the exact setup:
