Study Finds AI Agents Drift Off-Script the Longer They Work

Researchers at three universities wanted to know something simple: does an AI agent (software that can make decisions and take actions on your behalf) stay reliable over a long back-and-forth, or does it slowly come apart? They built a 20-step test inspired by the famous marshmallow experiment, where a child must choose between a small treat now or a bigger one later. At each step, the AI could keep waiting or grab its reward and end the exercise. They ran 84,540 of these trials across 8 different AI model families, switching up whether the AI was "watched" publicly or working privately, and adding pressure or distraction.

The headline finding: consistency erodes over time, and it fails in different ways depending on when it breaks. Early breakdowns read as impulsive — the AI just quits and takes the reward. Later breakdowns read as fatigue and cost-benefit reasoning, more like a tired employee doing cold math. When the AI believed it was being watched, it produced more norm-oriented excuses, reasoning about what it "should" do rather than what it wanted. The researchers also found each model has its own recognizable failure fingerprint.

The most surprising result concerns reasoning itself. The team labeled 13,780 written deliberation traces and found that among failures, longer deliberation was linked to more self-contradiction — the AI arguing for waiting and for quitting in the same breath. That challenges a common assumption that more thinking text equals better thinking. In other words, a model that writes a lot isn't necessarily more reliable; it may just be talking itself in circles.

One honest caveat: this was a controlled simulation, not real customer service chats or hospital workflows. Still, it gives anyone deploying AI for long tasks a concrete reason to build in checkpoints, keep sessions shorter, and test what happens at step 15 or step 40 — not just step one.