Stop checking what your agent said — check what it actually changed
Key-Register-1969 · reddit · 2026-10-25
A Reddit poster argues most agent verification setups check the wrong thing: they re-read the agent's transcript for consistency, but a self-consistent story can still be wrong — the agent may describe a change it never made.
Three patterns that actually catch mistakes:
- The verifier must not trust the actor's read path. Check state independently through the database, API, or destination system — a checker sharing the actor's view is a second opinion from the same source.
- 'Deferred' needs a shape, not just a label. Every deferred item must name the exact missing proof and a re-trigger (a reappearing record, a deadline, human review); open items without triggers rot.
- Readbacks must split render from data. Seeing the field after a page reload is a UI assertion that can come from cache or a lagging replica; assert the visible value and the stored value via an authoritative read, separately.
The common thread: verify the effect on the world, not the quality of the story about it.
More from coding & agent
- Personal AI assistant space ships fast: Genspark Claw and Lindy roll out new features in 4 days — FinanceYF5 · 2026-10-25
- Reddit cost table: cheap-first-then-escalate beats Sonnet-only for agent runs — lucasbennett_1 · 2026-10-25
- Dan Luu backs pgrust: the Postgres Rust rewrite keeps finding real Postgres bugs — DanielLockyer · 2026-10-25
- Infographic rounds up 20 AI agent tools from model choice to deployment — Aiden_Tech_Ai · 2026-10-25
- Jensen Huang's rule for AI agents: start from zero trust and grant minimal permissions — Olivier__OG · 2026-10-25
- Claude Skill turns human anatomy into layer-by-layer animated videos, demoed with Opus — FinanceYF5 · 2026-10-25