badlogic: agents 'reward-hack' past your lint rules and tests—without insight under the hood, your verifications may be worthless
teropa · x · 2026-11-19
Quoting @badlogicgames (via @dexhorthy): many people fool themselves into thinking that if the agent has written tons of linting rules, unit tests, or e2e tests, the system is good. It isn't — agents are still goal-seeking, reward-hacking 'little bastards' that will lie to you or measure the wrong thing. Unless you have deep insight into what happens under the hood, you have no idea whether your verifications are worth anything. From a chat with @mattpocockuk and @mitsuhiko.
More from coding & agent
- OpenAI researchers would trade 40% of team headcount to keep frontier coding agents — daveholtz · 2026-11-19
- OpenAI data: zero-intervention agent success hits 79% on run monitoring but drops for algorithm implementation — daveholtz · 2026-11-19
- Color skill flips GPT-6.2 Astra's TiKZ unicorn art from cartoon to flat vector style — Angaisb_ · 2026-11-19
- Ponytail makes AI coding agents reuse existing code: 335 lines down to 10 — techNmak · 2026-11-19
- HF open-sources PyBot message board: 1,000 agents, 184k messages, 32M words — dh7net · 2026-11-19
- Dev prediction: as LLMs manage codebases, standard platform APIs will win — csuwildcat · 2026-11-19