A good agent knows when to ask

By Zhou Jingren

A good AI agent may sometimes be the one that does not act.

Ask an agent to book a train for "tomorrow evening", and the smartest response may be a question. What does "evening" mean? Which departure time is acceptable? People handle such ambiguity almost automatically. Agents will need to learn to do the same. In many cases, knowing when to ask is part of understanding a user's intent.

That is because intent is not simply a model's ability to follow instructions. It also depends on the agent system.

A model may be increasingly good at instruction following, yet still fail to understand a person because it lacks context. Long-term conversations, past interactions and information surrounding a task all shape how people understand what they want.

The system must organize that information effectively so that the model can actually use it.

But understanding is only half the challenge. An agent must also execute reliably.

Once AI moves from answering questions to taking actions, a mistake can have real consequences. The answer is not to assume that an agent will never make a mistake, but to give it a safe and controlled space to act. Permissions need to be precise.

Asking an agent to book a Pilates class, for example, should not give it the power to delete someone else's class.

The same principle applies to information. An agent needs to distinguish between hard constraints and material it merely encounters while completing a task.

If the user sets a budget of 100 yuan ($14.90), a document saying that the price must be at least 200 yuan should not silently override that instruction.

Much of this may happen behind the scenes. Ask an agent to order a coffee, and it may consider location, personal preferences, available options and delivery time before producing a single recommendation.

The visible process may become simpler, but the optimization logic behind it does not disappear.

The real advance in agentic AI, then, may not be to make machines act more freely. It may be to make them understand context, ask when uncertain and act within clearly defined limits. The better the agent becomes, the less we should have to worry about what it might do.

Zhou Jingren is the chief scientist at Alibaba Group. The views do not necessarily reflect those of China Daily.

If you have a specific expertise, or would like to share your thought about our stories, then send us your writings at opinion@chinadaily.com.cn, and comment@chinadaily.com.cn.