Lauren Tan, a former Meta and Netflix engineer, now at Cursor, describes a system for autonomous AI agent operation in software development. She emphasizes that trust is the primary bottleneck in productivity; without a verification system, developers must manually check every action, which limits parallelization.
To achieve autonomy, Tan implemented AI Evals (verification). This allows agents to independently run code, capture CPU traces, and use iOS simulators to ensure correctness. She developed the Pstack plugin, which uses a feature map to provide agents with context on UI elements and navigation, reducing the need for manual guidance.
Tan also introduced the HAL skill, which prevents agents from guessing and forces them to read relevant code using sub-agents. To maintain these skills, she uses an eval playbook with coordinator agents and judge models for cross-verification, ensuring that skills remain effective as the product evolves.
Within Cursor, an agent named Benny automatically handles incoming bug reports by reproducing issues in a cloud instance of Cursor. For the Grokbot product, Tan created the 'Dune' architecture, which enforces strict CI checks, such as banning useEffect in React and prohibiting code comments to prevent clutter.
The system allows non-engineers, including designers and product managers, to safely submit code changes. Tan argues that while initial setup and token costs are high, the ROI is achieved through reduced hiring needs and increased team velocity. She mentions that the Grok 4.6 model has improved the price-performance ratio.
