Lauren

A Feature Map is a special file bundled with a skill that teaches the Agent how to navigate to every feature of the product from the user's perspective, including sub-features, keyboard shortcuts, and the data attributes of CDP-selected elements.

Lauren built two skills, 'how skill' and 'why skill,' enabling the harness to more rigorously examine the codebase and proactively investigate the codebase for follow-up details upon receiving a Slack message.

The trust ladder proposed by Lauren, ranked from highest to lowest effectiveness, is: the codebase itself, static analysis, rules/Bug Bot/skills, and style guides.

Lauren built a plugin named P-Stack, which is essentially a collection of skills that distills her years of software engineering workflows into corresponding playbooks and skills.

Lauren joined Cursor six months ago to work on optimizing performance issues in the Cursor Agent window.

Lauren advises against relying solely on style guides, noting that manual review is only suitable for catching what is missing, and recommends investing time in the earlier layers instead.

Lauren believes the biggest obstacle to transitioning from a small number of agents to a large number of agents is not compute or tooling, but trust.

Lauren believes that the core of an efficient Agent workflow is built around 'trust,' and engineers need to instill knowledge into the Agent to avoid becoming a bottleneck themselves.

Lauren previously worked on the React team and joined Cursor (later merged into SpaceX AI) about half a year ago.

Forcibly launching hundreds of sub-Agents in parallel in the absence of trust will generate a massive volume of junk PRs, regressions, and bugs.

The first validation skill Lauren built at Cursor is called 'Control,' and contains two key components: CLI and Feature Map.

P-Stack is a plugin built by Lauren that distills workflows such as debugging, feature development, and prototyping from her software engineering work into a skill library.

The 'trust ladder' proposed by Lauren contains five levels: the codebase itself, static analysis, rules/Bug Bot/skills, and style guides, with decreasing importance but varying degrees of enforceability.

When an Agent repeatedly makes the same mistake, rather than correcting it each time, it is better to add a lint rule or refactor the codebase so that the mistake is structurally impossible.

The Dune framework prohibits code comments because Agents often use comments as an excuse to cover up temporary patches rather than truly solving the problem.

Theo proposed a layered framework for handling agent errors: the first layer eliminates entire classes of errors at the architectural level, the second layer turns them into lint rules or tests, and only the third layer uses skills and rules as a fallback.

Lauren built a "control glass" skill that teaches agents to use the Chrome DevTools Protocol to run applications, capture traces, and achieve programmatic control.

Lauren now lets agents automatically merge PRs, and one morning woke up to find 20 PRs had already landed and were being reviewed directly on the main branch.

Cursor team engineer Lauren believes the true ceiling of AI collaboration lies not in model capabilities but in whether a trust system is established between humans and agents.

Lauren describes the mental model of Eval as "unit tests for agents," where a main coordinator agent formulates a scoring rubric and spawns sub-agents to conduct evaluations.