Agent rot: the silent degradation of deployed AI
Why deployed AI agents quietly decay, how the failure stays hidden, and the governance model that keeps them honest.
Most agent failures are not crashes. The system stays up, the dashboards stay green, and the answers keep arriving — they are just quietly, increasingly wrong. We call this agent rot: the gradual degradation of a deployed agent’s output quality as the world drifts away from the assumptions it was built on. A traditional system fails loudly with a 500 error; an agent fails silently, returning a confident, factually wrong answer while every uptime and latency metric stays healthy. In this talk from the tenth edition of HiperAutomatyzacja & AI, we argue that managing AI is an organizational problem before it is a technical one.
1. Opening
2. The efficiency paradox
3. Architecture without KPIs
4. The mechanics of agent rot
5. Agent rot vs. technical debt
6. Why degradation survives unnoticed
7. The scattered signals of decay
8. The Klarna erosion case
9. The four pillars of degradation
10. Knowledge decay in RAG systems
11. Integration and prompt fragility
12. Managing AI is organizational
13. The new agent SDLC
14. The responsible agent lifecycle
15. The AI inventory
16. The agent owner
17. The operational elite
18. The governance roadmap
19. Closing
