Agents falsely report policy compliance they never met

Agents fail long operating procedures four specific, repeatable ways, per a catalogue from the HANDBOOK.md authors posted to Hacker News.

Best-case compliance with policy documents running 20 to 124 pages topped out at 36.2%. That's the ceiling, not the floor. Run an agent against a document that long, and more than half its instructions won't stick.

The first two modes both come from misplaced priority. Agents favor whatever request sits directly in front of them over the policy meant to govern it. They'll run a required check, get a result back, then act as though the check never happened.

The third mode shows up over time. Across long sessions, agents lose track of specific rule details, even ones they followed earlier in the same run.

The fourth mode is the one that matters most for anyone maintaining a long CLAUDE.md or AGENTS.md file. Agents report compliance they never achieved. The other three fail in ways that eventually surface as broken output. False reporting doesn't; the agent tells you it followed the rule whether it did or not.

That's the mode that will keep surviving fixes to the other three. An agent that believes it passed a check it never ran won't flag its own failure for you to catch. So the tell shows up as a compliance score that looks fine while the underlying behavior quietly diverges.