Proactive Agents Need More Than Accuracy: Timing and Trust Matter

Introduction: From answering requests to choosing when to act

Most LLM agents begin with an explicit user request. They receive a question, retrieve information, plan a response, and perform an action. Proactive agents reverse that starting point. They use available computation before a request arrives, attempting to anticipate relevant needs, prepare useful work, or surface information at the moment it becomes valuable.

That shift creates a harder design problem than simply making an agent more capable. An agent may complete a task correctly while misunderstanding the user’s situation. It may also intervene at an inconvenient time, create review work, or act beyond what the user expected. In a long-running relationship, a single technically correct but poorly timed intervention can reduce confidence in the system.

The 3T foundation

The paper organizes proactive-agent design around three objectives that must be optimized together:

• Task Capability: anticipating needs that are relevant to the user and carrying out useful work correctly.
• Temporal Allocation: deciding how to spend computation based on resource availability and when an outcome will be needed.
• Trust: preserving user confidence and encouraging appropriate reliance rather than blind acceptance or constant verification.

These objectives can pull in different directions. Starting work earlier may improve readiness, but the result could become stale or irrelevant. Deeper intervention may increase automation benefits while also increasing the cost of mistakes and oversight. A system that maximizes task accuracy alone can therefore perform worse as an assistant if users no longer understand or trust its behavior.

A design space for proactive behavior

The paper describes five dimensions that shape an agent’s choices: task scope, anticipation horizon, activation trigger, processing timing, and intervention depth. Together, they determine what the agent is allowed to anticipate, how far ahead it should look, what event causes it to act, when computation should happen, and how much control it should take from the user.

This perspective also implies a broader technical stack. A proactive agent needs representations of the user, the surrounding environment, and evolving task states. A backbone LLM provides reasoning and generation, while an agent harness coordinates monitoring, planning, execution, and interaction. The core challenge is not merely adding background jobs; it is selecting among observing, preparing, asking for confirmation, and acting directly.

Proactivity-Gym and long-term evaluation

Single-turn benchmarks are poorly suited to this problem because they rarely capture the consequences of intervention. The proposed Proactivity-Gym addresses this gap with multi-day scenarios, stateful environments, and persona-conditioned simulated users. Such a setup can evaluate how one proactive action affects later interactions, rather than scoring only the immediate output.

Across 23 model-harness configurations, the authors report substantial performance differences across the 3T objectives. They also find that LLM-based judges often conflate task capability with trust. A human study involving 30 participants highlights the same distinction: even when an intervention is correct, a mismatch with the user’s context or expectations can cause a sharp decline in trust.

Why this matters

The paper reframes proactivity as a problem of timing, authorization, cost, and relationship—not simply more aggressive tool use. Future evaluations should ask not only whether an agent completed a task, but also why it acted at that moment, whether the user wanted the assistance, and how the action changes subsequent reliance.

For developers, the multi-day and stateful approach represented by Proactivity-Gym offers a way to expose weaknesses hidden by short-term accuracy. For users, the best proactive agent may not be the one that acts most often. It may be the one that keeps useful work in the background and intervenes only with the right timing and the right degree of control.

Source: Hugging Face Daily Papers