Proactive Agents: When They Judge Wrong, No One Knows
A proactive agent and an ordinary agent differ by only one step: after receiving a trigger, it first decides whether it should act, when to act, and how far to go. But no one is watching this step, and if it gets it wrong, no one knows. The key to evaluation is to make every 'non-action' visible.
Muse, Grok Bot and other proactive agents have recently become a hot topic in the AI world. They don't need you to speak; they watch your calendar, inbox, and project progress on their own, and come to you when they think something matters.
The difference from an ordinary agent is really just one step.
An ordinary agent executes as soon as it receives a trigger. A proactive agent, upon receiving a trigger, first decides: is this thing worth doing, should it be done now or wait a while, and how far should it go? For example, if a flight is delayed by two hours, it can directly remind you, wait for the airline to rebook, help you find a new flight, or even book it for you. When an email comes in, the trigger is the same; the judgment process determines whether it is a helper or a nuisance.
This judgment step is a link no one watches.
Every answer from an ordinary agent is watched by someone, and a wrong answer is immediately discovered. A proactive agent most of the time judges "not worth bothering with" and quietly moves on; if that judgment is wrong, no one knows. This is the hardest part of evaluating it: you cannot evaluate it only through "what it did," because "doing nothing" is most of its output.
Annabell from Langfuse wrote a long article specifically about this. Her solution is: record the judgment itself.
For every event that reaches a decision step, record four things:
- what event triggered it
- which context it pulled from memory and connected apps
- the final decision: act, wait, or stay silent
- what the reason was
The key is that records of those runs where "nothing was done" must also be kept. A decision not to act produces no output; unless it is proactively put into a review queue, no one will ever see it. Silence is also a result, and it is often the correct result—provided you can go back and check it.
The evaluation angle can also be changed.
Every judgment is really answering a question: should action be taken on this matter? This is essentially a classification task. Look at it from two sides: when action should be taken, was anything missed; when action should not be taken, did it act unnecessarily? And don't set just one threshold. Reminders that can be ignored casually can be more permissive, while operations like buying things, which cost money and are hard to undo, need a much stricter threshold. The original article gives an example: a user reported that Instinct, when shopping and encountering a login wall, reset the user's password—rather than asking the user. Over-action is not just annoying; it also has security consequences.
The dataset also needs a new way of thinking. The evaluation set for an ordinary chat agent is "question + standard answer." The evaluation set for a proactive agent is "moment + expected decision":
- Input: trigger event + snapshot of user memory and instructions. Without this snapshot, judgment is impossible—the same flight delay should lead to silence for one person and rebooking for another.
- Expected output: act, wait, or stay silent; how far to go; plus a brief reason.
Sample sources can be user feedback (ignoring, undoing, "why didn't you tell me"), random sampling (including runs where nothing happened), and pairs of the same trigger across different users.
When a judgment is wrong, it is important to distinguish whether the problem is the judgment or the memory—if the memory is wrong, the judgment will be wrong as well.
Proactive agents' capabilities, tools, memory, and triggers—everyone has these components. What really determines whether one is good or bad is that one function: whether this moment is worth acting on, when to act, and how far to go. And this function runs where no one can see it. Only by making it visible can it be measured and improved.
A netizen summed it up even more plainly: whether a proactive agent is useful comes down to one sentence—bother when it should bother, and don't bother when it shouldn't.
Original article: Evaluating proactive agents
发布时间: 2026-11-20 19:33