X Discussion: Grok Bot Engineer poteto Calls Adversarial Agent Code Reviews "Mostly a Waste of Tokens" and Argues for Empirical Plus Formal Verification Instead

Background and industry pain points

Coding agents now write code faster than people can review it, so review has become the new bottleneck. A common answer is multi-model adversarial review: spin up several models to attack a diff and rely on model diversity to catch bugs. The cost is tokens, since every review sends the diff and its context to several strong models. On November 3, poteto (lauren) posted what she called "maybe a hot take": she now considers adversarial reviews mostly a waste of tokens. Per her X and GitHub bios, she works on Grok Bot at SpaceXSI, is on the React Compiler core team, and previously worked at Cursor, Meta and Netflix. She maintains pstack, an engineering skill pack for Cursor and Claude Code.

Architecture highlights and mechanisms

The skill she named is pstack's interrogate. Per the public SKILL.md in GitHub cursor/plugins, it spawns one read-only reviewer per configured model (defaults Claude Opus 5.5 xhigh and Grok 4.7 xhigh fast) to adversarially review the branch diff, then the lead agent sorts findings into Act on, Consider, Noted and Dismissed without auto-applying fixes. Her alternative is empirical plus formal verification: prove problems with real runs, tests and repro scripts, and prove properties hold with formal methods, rather than asking models for opinions about the code. In her words, she has found very little value in code review in the abstract, because agents tend to get lost in nitpicks and YAGNIs (you aren't gonna need it).

Benchmarks and measured comparisons

This is a personal judgment from experience; the post gives no comparison data or benchmarks. The only numbers are its reach: about 906K views, 9,227 likes, 5,538 bookmarks, 668 replies, 356 reposts and 210 quotes at fetch time, and the high bookmark count suggests many developers saved it for reference. She says pstack now uses interrogate less, but at fetch time the public pstack folder in cursor/plugins had no commits after October 27, so the change is not visible there yet.

Developer adoption and getting started

If your agent pipeline runs multi-model review, first measure how many of its comments actually get acted on and how many are style points or "might need it later" suggestions before deciding to keep it. One direction to borrow is evidence-required review: every finding must come with a failing test, a repro command or output, and anything without run evidence is dropped; add property tests or formal specs for critical logic. To try pstack, run /add-plugin pstack in Cursor, or install it from the cursor/plugins marketplace in Claude Code. Keep in mind this is one engineer's view, and teams should decide based on their own defect data.