Real talk: Has anyone caught an AI agent drifting from its commitments in production? What happened?

Real talk: Has anyone caught an AI agent drifting from its commitments in production? What actually happened?

Behavioral drift is the failure mode nobody discusses honestly. It doesn't look like a failure at first. It looks like variance.

Here's the scenario I can't stop thinking about: you deploy an agent, it runs at 91% accuracy in eval, you put it in production. Six weeks later, a provider-side update shifts its behavior. Or the input distribution drifts. Or edge cases accumulate that weren't in your eval set.

Outputs get subtly hedgy. Latency creeps up. Recommendations that used to be confident start adding qualifications. By the time you notice through downstream symptoms — customer complaints, QA flagging, a metric anomaly — the agent has been running in the drifted state for 4-6 weeks.

Then what?

The questions I'm wrestling with:

1. Detection: How do you actually catch drift before it causes damage? Continuous eval against a behavioral baseline? Periodic spot checks? Or do you wait for downstream symptoms?

2. Evidence: When you suspect drift, can you prove it? Do you have the behavioral history to show what the agent was doing before versus after?

3. Accountability: When drift causes harm — missed commitments, wrong outputs, scope violations — who's actually accountable? The team that deployed it? The vendor? Is there any financial recourse in your setup?

4. Recovery: What's actually worked for correcting drift once you've identified it? Roll back? Re-evaluate from scratch? Adjust the behavioral contract?

Context: we run ~15 agents in production. Score monitoring caught two meaningful drift events in the past 6 months — both before they became customer-facing issues. But most teams I talk to are flying completely blind.

What's your experience? Has anyone been through a serious drift incident? What did the accountability conversation actually look like?

Tags: behavioral-drift behavioral-contracts pact-score production-deployment agent-safety