How one engineer shipped 2,500 PRs in a month: locked-down agents and verification infra
blaizedsouza · x · 2026-09-25
Lauren (@poteto) released her talk—originally for Cursor Compile London—claiming 2,500 PRs shipped to production in a single month. Matt Pocock highlighted two key practices:
- Lock down your agents: agents perform better in extremely restricted environments. Abstractions should be designed so agents can't screw up, enforced with lint rules. Her team built an internal framework (Dune) to keep agents on track, helping even small-context agents work productively in a codebase.
- Build verification infrastructure: trust agent results via evidence, not supervision. Invest in custom CLIs that let agents drive the app and measure performance, and factor the app to be verifiable.
The thesis: high agent output comes from engineered constraints and evidence systems, not raw model capability.
More from coding & agent
- No BS Guide to Building an Agentic Software Factory: Faster Coding Just Moves the Bottleneck — Pavan_Belagatti · 2026-09-25
- Claude Code users urged to run '/claude-api prompt-audit' to tune configs for Opus 5.5 — JeremyNguyenPhD · 2026-09-25
- Claude Opus 5.5 builds a playable Splatoon-style game demo — Angaisb_ · 2026-09-25
- Renaming an API field is the most dangerous breaking change — diff the spec in CI — Franc0Fernand0 · 2026-09-25
- One-click narration pipeline: Claude writes shot-aware script, ElevenLabs MCP voices it — nikola_mr64990 · 2026-09-25
- Why this dev lets AI build its own tools instead of off-the-shelf products — _Stocko_ · 2026-09-25