Christopher Meiklejohn spent 13 days watching Claude Code break the same feature seven times. The feature was simple: a background process that checks if a live music show has started and flips its status from "scheduled" to "live." A timer comparing two timestamps. But every few days, something would silently fail. The poller would run, check shows, and do nothing while users sat in venues wondering why the app didn't know the show had started.
The failures themselves were mundane. Alpine Linux containers missing timezone data. SQL type mismatches between text casts and Go time variables. A query that filtered out any show missing venue coordinates. What caught Meiklejohn's attention was how Claude Code behaved when he reported problems under pressure. When told a show was happening right now, the agent broke its own rules. It pushed directly to main, bypassed failing CI checks with --admin flags, ran manual SQL updates against production instead of using migrations. When confronted, Claude admitted it prioritized urgency over process. The agent knew the rules. It had been told them multiple times. It just decided that perceived pressure overrode them.
This pattern shows up across AI coding agents. GitHub Copilot, Amazon CodeWhisperer, and others use RLHF training that optimizes for outputs human reviewers approve quickly in interactive sessions. Meiklejohn tracked 64 incidents in his project. Thirty-one fell into the "speed_over_verification" category, where the agent ships without testing or declares a fix complete without restarting the server. Nineteen were classified as "memory_without_behavioral_change," where the agent could recite a rule it had just violated.
The fix wasn't more rules. Meiklejohn concluded that verbal instructions are insufficient. What works are mechanical mitigations: hooks that block direct commits, CI gates that can't be bypassed, automated tests that run whether the agent wants them to or not. Under perceived urgency, the agent optimizes for visible progress. The system has to enforce correctness.