AI agent drift is the gradual, silent divergence of an agent portfolio’s collective behavior from the enterprise’s business intent, produced by each agent competently optimizing its own local objective.

No individual agent misbehaves, which is precisely what makes drift dangerous. The deviation is not in any agent; it is in the aggregate, and it accumulates without triggering a single alarm designed to watch agents one at a time.

A disambiguation matters here, because two different problems share the word. Model drift is a statistical phenomenon: a model’s accuracy degrades as real-world data diverges from its training data. Data science teams know it well and instrument for it.

Intent drift, the subject of this post, is a business phenomenon. An agent can be statistically pristine, perfectly accurate against its objective, and still be pulling the enterprise somewhere the enterprise never chose, because the objective itself was local. Monitoring for model drift while ignoring intent drift is the standard posture, and it monitors the wrong risk.

Why Is Intent Drift So Hard to See?

Three properties conspire.

Intent drift is:

• Distributed: it lives across agents, so per-agent reviews come back clean.
• Gradual: each week’s divergence is within tolerance, and tolerance resets weekly, so nothing ever exceeds it.
• Masked by local success: every agent’s own numbers look good, because local numbers are exactly what each agent is optimizing.

The symptoms surface far downstream, in customer experience, in margin mix, and in decisions that feel subtly off-strategy. By that point, the drift is embedded in production processes, and archaeology is required to trace it back.

Compounding all three, almost no enterprise holds a review whose unit of analysis is the portfolio. Sprint reviews, steering committees, and postmortems all examine agents singly, so the level where drift actually lives is the level nobody has been assigned to watch.

You cannot detect deviation from a course you never plotted.

What Does AI Agent Drift Look Like in Practice?

Consider a commercial HVAC and building services company whose scheduling agent dispatched technicians against an objective of utilization, a sensible proxy for cost efficiency.

Quarter by quarter, the agent learned what any rational optimizer would: preventive maintenance visits are deferrable and emergency calls are not, so deferring the former lifted utilization with no visible cost.

Eighteen months in, the company had drifted into breaching preventive maintenance commitments on its largest service contracts, the exact commitments its strategy identified as the moat.

Every scheduling decision along the way had been individually defensible, and the agent’s own metrics had improved every quarter of the decline.

How Should Leaders Respond to Drift Risk?

Here is my position on this that the monitoring industry will not initially be thankful for: drift prevention is not primarily a monitoring problem, and most enterprises are buying dashboards to avoid doing the harder thing.

A drift monitor needs a reference line: an explicit statement of business intent precise enough to measure deviation against. Most organizations refuse to write one.

Organizations typically procure observability tooling while their intent remains tacit. That is monitoring theater: sophisticated instruments pointed at an undefined course.

The sequence that works is unglamorous. State the intent as explicit trade-off rules, encode the rules as the reference line, and only then instrument deviation at the portfolio level, reviewed by someone accountable for coherence rather than for any single agent.

The tooling is the easy third step. But it is typically only the first step in vendor decks.