When agents replace humans on tasks, undocumented organizational decisions become permanent liabilities – agents extend existing patterns without knowing whether they are intentional or historical accidents.
Key Takeaways
Agents have only the context window; when a session ends, unwritten context is gone, not deferred – every session starts from zero.
Intent debt propagates silently: an agent sees every customer integration has its own service and creates another, unaware the pattern came from a scaling constraint that no longer exists.
ADRs add friction between “this pattern exists” and “this pattern should be extended” by recording what was considered and why alternatives lost, not just what shipped.
Specs prevent agents from silently removing load-bearing logic – retry code that looks redundant may be a required idempotency guarantee the code itself never explains.
Playbooks give agents the cause, not just the symptom: restarting a pod clears a memory spike; the playbook says check the upstream rate limiter configuration first.
