The Agent Recovers; the Bad Instruction Survives

Repository instruction files are read on every coding-agent session, so stale paths and procedures can accumulate interest like documentation debt. Researchers tested three real-world stale instructions—a test command, code location and component procedure—across 22 maintenance tasks, three instruction variants and two models, yielding 732 trajectories. Agents followed relevant stale guidance even when correct information existed elsewhere. One stale test command cost one model about two extra tool calls per run, while the other cases showed no detectable difference. Agents generally recovered from immediate failures, but none corrected the underlying instruction file. Successful task completion can therefore conceal recurring operational waste.