Cursor engineer Lauren Tan on letting AI agents earn trust to merge code
xiaohu · x · 2026-09-08
Lauren Tan, a Cursor engineer (formerly Meta's React compiler and Netflix tech lead/EM), shared in a Maven talk how she went from watching an AI agent's every step to letting it merge code autonomously. She managed the agent like a new colleague: it can code, but doesn't know the product or team conventions.
She built three conditions:
- Let the agent run the app itself and verify its results;
- Encode product knowledge and workflows via feature maps and skills, evaluated with real tasks;
- Turn recurring mistakes into automated checks so the right way is the easy way.
Once stable, cloud agents and other teammates can reuse these setups. The talk starts from a stubborn performance bug and answers a concrete question: what does an agent that writes patches but always needs review actually lack — and how trust gets built incrementally.
Related event: Cursor Engineer Shares How to Trust Coding Agents with Auto-Merge(2 posts)→
More from coding & agent
- POV: asking Gemini to fill a Google Sheet with data from Gmail — omooretweets · 2026-09-08
- Codex CLI tip: Ctrl-O copies the last response as Markdown, /copy grabs just the relevant part — gabrielchua · 2026-09-08
- herdr 0.9.0 ships the most-wanted feature: one client to control agents across all your machines — ycombinator · 2026-09-08
- Hands-on with Grok Bot's cloud agent team: hiring an AI staff without a server closet — RachelVT42 · 2026-09-08
- Should Agent Skills scripts be prewritten or generated on the fly? — dotey · 2026-09-08
- Embedflow: zero-downtime embedding model migration without 108-day backfills — Potential_Low_1183 · 2026-09-08