How Perplexity verifies Brain's skill edits: engineer approval vs automated checks
perplexity_ai · x · 2026-11-20
Perplexity tiers verification by blast radius:
- Built-in skill revisions affecting all users must pass every test scenario, fail at least one on the original, and get engineer approval
- Personal skill proposals pass deterministic checks, with an LLM judge auditing a sample
- The diagram breaks proposal checks into exact edit, evidence, baseline match, and eligibility
Related event: Perplexity Rolls Out Auto-Configured Personal Skills to All US Users(4 posts)→
More from coding & agent
- Cognition launches Devin Bots: Exa runs training and CI with agent teams — marvinvonhagen · 2026-11-20
- Eval-driven development: pick your model, retrieval, and agent design by testing, not guessing — blaizedsouza · 2026-11-20
- Open-source Harness Engineering course: 14 lectures, 8 projects on making coding agents reliable — blaizedsouza · 2026-11-20
- Google Cloud outlines 5 research-backed patterns for agent orchestration — rseroter · 2026-11-20
- Hud turns 'It looked good to me' — AI code review's most dangerous words — into a musical — iamfakhrealam · 2026-11-20
- Sentry CEO: Claude's memory seems to just inject problems into context — zeeg · 2026-11-20