Skip to content- Sources: primary
- Summary: At AI Engineer, Lauren Tan said the Grok Bot client team shipped 11,000 pull requests to production last month from about 227 contributors, with reverts steady around 0.5% and startup and typing performance improving. The team bans patterns agents misuse, such as React useEffect in product code, imports between Electron main and renderer directories, and 99% of code comments, enforcing each with lint rules and a dependency graph check. Review now centers on verification artifacts from runtime skills with a feature map, TLA+ model checking of notification flows, Lean proofs for a virtualization library, and property-based UI tests, and a release-time agent fuzzes the build and auto-merges its fix PRs.
- Why it matters: It gives a concrete order of investment before unattended merging: deterministic constraints in code first, then runtime and formal verification, with prose rules last.
- Background: Teams adopting coding agents often add instructions to files like AGENTS.md when an agent errs, but prose rules are not enforced and agents copy whatever patterns already exist in a codebase. Tan's approach converts each correction into a deterministic check so review load shrinks over time. TLA+ is a specification language whose model checker explores all states of a design, and Lean is a proof assistant for machine-checked proofs.