The agent method behind 2,500 PRs a month: don't copy that prompt yet
Lauren Tan's pstack and Chonsy's Grok Bot tutorial both discuss agent teams. What really determines whether they are usable is permissions, verification, and the evidence chain. The Opus 5.5 part is actually the easiest to misread.
Lauren Tan shipped 2,500 PRs to production in a month, and that number has been reshared many times on X. She is a SpaceXAI engineer, and the method she shared is called pstack. The core actions are not complicated: let the agent start the app itself, run the checks related to the change, save evidence, and then write reusable verification instructions into the repository so the next agent can use them.
Chonsy then wrote a tutorial that applied Lauren's method to Grok Bot, Cursor cloud agents, and GitHub. The tutorial title includes Opus 5.5, but there is an easy-to-misread point here. Chonsy states in the article that Grok Bot selects the backend model itself; you cannot force Opus 5.5, nor can you inspect which model was used behind a given request. If you want to explicitly choose a model, you need to confirm it in the Cursor cloud agent configuration. Model selection and run results must be recorded separately. Treating the title as proof of model capability will lead you astray.
The team in the tutorial has four roles. Chief takes requirements and records the owner and blockers; Lead breaks the issue into a plan with acceptance criteria; Builder implements on an independent branch; Reviewer uses the original requirements, the PR's exact head SHA, and Builder's evidence to perform an independent check. The four bots are placed in a group chat, and handoffs must clearly state the issue, current owner, commit, evidence, and next step. This constraint matters more than the role names. With ten agents working in parallel, if nobody knows which commit was verified by whom, in the end you just have ten PRs waiting for human review.
Lauren's permissions prompt covers request.bind, authority.resolve, action.inspect, action.route, context.sync, worker.cycle, proof.match, check.record, result.route, repository.maintain, return to_thread, and autonomy.expand. There are many names, but in practice they govern four things: the request source and existing authorization must be bound; actions must inspect the target and impact; before execution, the baseline must be reproduced and fixed; results must include the commit, worktree, command, expected, actual, exit code, and evidence path. action.route classifies actions into ALLOW, CLARIFY, and ESCALATE. One line in authority.resolve is critical: understanding != authorization, passing_checks != release_approval. Understanding the task does not mean you have permission, and passing checks does not mean you can release.
The verification part lives in a skill in the repository. Chonsy's example uses the Cypress Real-World App. First, have Chief confirm that dependencies can be installed, the frontend and API can be started, and yarn types and the auth tests can run. Then use pstack's create-verification-skill to generate .cursor/skills/verify-rwa/SKILL.md, clearly documenting launch, doctor, drive, evidence, and cleanup. The feature map starts with authentication and records the entry point, selectors, actions, expected results, and failure conditions. When Reviewers check, they must check out the PR's exact head SHA, run the relevant checks, and keep logs and screenshots. If there is a new commit, the old verdict is invalid.
Someone in the replies said that after using poteto's skills for a while, the part most worth copying is the verification results. The PR includes a checked-off test plan and adds a line for production-environment testing; when the verify skill cannot cover the CI workflow, the PR must state this and fill the gap with fixtures plus a dry run. This point is more useful than only discussing models. Verification results cannot just say tests passed; they must let the next person reproduce them.
CI is also connected to this process. The example repository adds an auth-verification job that starts yarn dev on pull_request, waits for ports 3000 and 3001, runs types and the auth spec, and uploads Cypress screenshots and videos. Only after it is configured as a required status check does it become a merge gate. The Reviewer's PASS is another report, and before merging you still need to align the reviewed SHA with the current head. Any new commit must be checked again.
People in the replies also worried about permissions. One user asked: are you really going to hand this whole setup directly to agents without checking everything first? That worry is not unnecessary. No matter how detailed the permissions prompt is, pasting it into a chat box will not automatically turn it into governance. It needs repository rules, tool permissions, approval conditions, and execution records to work together. The autonomy.expand section is written with restraint: first have repeatable verification, then expand delegation; define permissions, isolation, and stop conditions; keep decision logs and rollback options. Without these, an agent team merely lengthens the human review queue.
The reusable lesson is in boundaries and evidence. You can have four roles, or three; models can be swapped; prompts can be trimmed. Let agents work within inspectable authorization boundaries, and leave the source, impact, and evidence of every action in the repository and PR; only then can this setup support parallelism. If you want to try it, first get environment verification working for one repository, then generate a verify skill for one feature, and then let one issue go through Lead, Builder, and Reviewer. Do not rush to start ten agents. First see whether the evidence from the first PR lets you ask three fewer questions.
发布时间: 2026-11-17 21:34