How a Grok/Cursor principal engineer ships ~2,000 PRs a month with coding agents
dumpshoot · reddit · 2026-10-27
Lauren Tan ("poteto"), Principal Engineer on Grok Bot and Cursor, explains how her personal collection of skills lets her ship 2,000 PRs a month to production — 2,462 in August alone. PR counts and LOC don't matter by themselves, she says, but when agents raise output while keeping quality, scale starts to matter.
Part 1: agents must verify their own work
- Verification is core infrastructure: agents need to run the app, check results, and loop until requirements are met; humans supervising every step become the bottleneck. Getting this right can boost team productivity 100–1,000x (her estimate, not benchmarked)
- Give agents verification tools per stack: Chrome DevTools Protocol for web/Electron, simulators for iOS; build a CLI if needed (sessions, commands, screenshots, accessibility tree, performance traces, logs)
- pstack's "Build the Lever" principle: composable commands, clear instructions, JSON output, error messages that state next steps, --dry-run for risky ops
- Feature Map as "condensed memory": what features exist, their purpose, entry points and shortcuts — maintained daily via /maintain-verification-skill, or agents verify against stale info
- For parallel agents she avoids Git worktrees (disk/CPU heavy) in favor of cloud agents with their own machines; Cursor snapshots the initial setup for faster subsequent runs
Part 2: don't expect agents to decode vague requests
Two recurring failure modes: unclear initial requests, and missing codebase context. Output quality hinges on context preparation, not the model. One technique: indirect prompting — have the agent restate the problem in its own words (article continues beyond excerpt).
More from coding & agent
- ComfyUI prompt generator GOAT'd adds cross-run context memory and structured JSON prompts — ninjaGurung · 2026-10-27
- 7 Claude prompts to build a monetized AI YouTube channel in 90 days — Aiden_Tech_Ai · 2026-10-27
- Asana survey: 46% of multi-agent users say managing agents eats 10%+ of their work time — ombre-ai · 2026-10-27
- Matt Pocock names the sweet spot: tasks where agents have full data and can self-verify — mattpocockuk · 2026-10-27
- xAI releases free Grok Bot course: 18 real-team playbooks, ~3 hours, covering sales to engineering — victor_explore · 2026-10-27
- Building agent infrastructure for specialized domains is the hard part, not the models — Electronic_Sort_6627 · 2026-10-27