Lauren Tan's pstack workflow: how one engineer shipped 2,500 PRs a month with agent teams

garrytan · x · 2026-11-01

Garry Tan amplified Lauren Tan's demo of pstack, an agent execution workflow, complete with a detailed execution map:

Routing: /poteto-mode interprets intent into goal/constraints/observablesuccess and selects a task playbook (bug fix, feature, refactor, performance), with /how, /why, /architect skills supplying context and design; /swarm and /arena split work or compare approaches.
Verification contract: /create-verification-skill builds a repo-specific verifier: launch → doctor → drive → evidence → cleanup. Proof is matched to the change — UI completes the affected user journey, CLI captures command + exit status, storage reads back written values, performance compares profiles under equivalent conditions.
Reviewable handoff: failed checks bounce back for repair; unavailable checks stay listed under unverifiedchecks; delivery includes testedcommit, diff, commands + results, evidencepaths.
Preventing repeats: recurring corrections become architectural constraints, type/lint/CI blocks, tests, or documented rules.

An accompanying article describes building an agent team with Grok Bot + Opus 5.5, where a lead bot splits work, delegates scoped tasks and collects PRs with verification results attached — reportedly 2,500 PRs shipped to production in a single month. The core argument: manual review doesn't scale to ten parallel agents; verification must be automated.

Original post →

Lauren Tan's pstack workflow: how one engineer shipped 2,500 PRs a month with agent teams

More from coding & agent