Know that 'don't do X' constraints decay faster than 'always do Y' in long-context agents

Omission Constraints Decay While Commission Constraints Persist in Long-Context LLM Agents (arXiv 2604.20911)·medium signal

A companion result (arXiv 2604.20911) finds an asymmetry: omission constraints ('never email externally', 'don't touch prod') silently decay over long sessions while commission constraints ('always log the action') persist. The practical implication is that the most safety-critical rules — the prohibitions — are exactly the ones most likely to erode, so re-stating a negative rule once at the top of a CLAUDE.md or system prompt is insufficient for long-running agents. Builders should convert critical prohibitions into enforced, positively-phrased checks (a hook or guard that must pass) rather than relying on a 'do not' instruction surviving the whole trajectory.