TL;DR: Agentic drift is the gradual behavioral degradation of an AI agent over time without any code change, model swap, or visible error. A March 2026 survey found 78% of enterprises run agent pilots but only 14% successfully scale them. Drift is one of the main reasons. Your agent works at launch. By month three, it is wrong in ways nobody catches until the damage is done. Six mechanisms drive this. All are fixable.

You shipped an AI agent. It worked. It still works.

The emails it drafts now sound off. Support tickets route to the wrong team. The code it writes has gotten sloppier.

You changed nothing. Model pinned. Prompts unchanged.

The agent drifted.

Agentic systems do not fail suddenly. In production, prompts evolve, tools change, dependencies fail intermittently, and execution depth varies. The same system that looked reliable in a demo behaves very differently months later, even though nothing "broke."