Lauren Tan: 2,000 PRs a month via Trust Engineering

Tan frames the problem as one many teams will recognize: once agents write code, people end up reviewing every line and issuing small corrections in chat, which she says is far from real scalability. Her answer is not to wait for better models but to design the environment so that agents can work on their own.

Her five-level Trust Hierarchy explains the reasoning. Physical constraints in the codebase and architecture scale most automatically, followed by static analysis such as lint and CI, then rules and playbooks, then natural-language style guides, and finally human review, which depends on people and barely scales at all. In the same spirit, she describes herself as a gardener who pulls out workarounds, since large language models tend to imitate whatever code already sits in their context, and a single patch can spread like a virus.

The supporting infrastructure she describes includes ControlGlass, a verification CLI built on the Chrome DevTools Protocol that lets an agent launch an app, collect traces and analyze heap snapshots on its own, plus a Feature Map that links vague bug reports to exact DOM elements and code locations, and Pystack, a set of playbooks drawn from senior engineers' workflows. Outermost sits the event-driven layer: Grok Bot Routines watch Sentry alerts and Slack threads, then launch cloud agents that reproduce bugs and open fix PRs with performance statistics. Whether this model travels beyond her own team is likely to depend on how much of that tooling a typical organization is willing to build, since the whole approach rests on that groundwork rather than on any single model.