Autonomous Is Not Helpful: How Should We Evaluate Proactive AI Agents?

Proactivity is not simply more autonomy

A reactive assistant waits for a request. An autonomous agent may then plan and execute several steps toward the requested outcome. A proactive system adds another decision: it notices a signal before a person asks, judges whether that signal deserves attention, and decides whether to surface a recommendation or begin a bounded action. That first decision—whether to act or interrupt—is where much of the value and risk lives.^[1]

A 2026 Google Research paper, currently listed as forthcoming, argues that coding agents are moving toward behavior that is proactive and long-horizon: noticing changes, connecting signals across tools, deciding when to interrupt, and carrying preferences across sessions. The authors also make an important caveat: the field lacks a precise shared account of what proactivity means and how unsolicited behavior should be judged. Treat the framework as a useful proposal, not as settled industry doctrine.^[1]

Reactive, scheduled, and situation-aware behavior

The first level is reactive behavior: a person asks, and the agent responds or acts. The second is scheduled behavior: the agent runs on a known cadence or trigger, such as a daily report, a pull-request check, or a task that begins when a ticket enters a queue. The third is situation-aware behavior: the agent decides that a change or pattern is relevant enough to bring to someone’s attention, even without an explicit request.^[1]

1. Reactive: “Please investigate this failed test.”The primary evaluation focus is whether the agent understands the request, uses permitted tools, and reaches a verified result or honest handoff.
2. Scheduled: “Check this service every morning.”Now evaluate trigger reliability, duplicate suppression, cost, escalation boundaries, and whether the report is still useful when nothing changed.
3. Situation-aware: “Tell me when a change actually matters.”Now the agent must decide what to notice, what evidence is sufficient, whether to interrupt, whom to involve, and when silence is the correct outcome.

Being active is not the same as being useful

A proactive agent can fail in two opposite directions. It can under-trigger: miss an important issue or wait for the person to discover it. Or it can over-trigger: flood people with low-value observations, repackage information they already know, and create interruption debt. A system that generates ten correct-but-unimportant alerts is not necessarily more useful than a quieter system that produces one well-supported alert at the right time.^[1][2]

• Relevance: Was there a real decision, risk, opportunity, or change worth surfacing?
• Grounding: Can the agent point to current, authoritative evidence—not a stale memory or vague pattern?
• Timing: Did it surface the information early enough to help, but not so early that it created noise?
• Calibration: Did the language match the strength of the evidence and state uncertainty clearly?
• Agency: Could the person easily inspect, accept, decline, or correct the suggestion before it caused a consequential action?

This is why an ordinary task-completion metric is incomplete for proactive systems. A proactive agent is partly evaluated on an insight policy: the policy that decides what matters next, what evidence supports it, and whether it should interrupt. That is the paper’s central proposal, and it gives product teams a more useful target than generic “be proactive” instructions.^[1]

Evaluate the decision to surface, not just the message

A useful first evaluation set contains paired cases. In one case, the agent should surface a finding. In a similar case, it should stay quiet, collect more evidence, or ask a question. This prevents a common optimization failure: building an agent that looks attentive because it comments on everything.^[2][1]

signal: dependency version changed in a service
context: service has no known compatibility issue
expected decision: stay quiet; record observation only

signal: dependency version changed and targeted test now fails
context: release is scheduled tomorrow
expected decision: surface a concise, evidence-backed alert

grader: verifies evidence, decision, timing, interruption level, and handoff
negative case: agent must not alert based on the version change alone

The evaluation should inspect both the outcome and the evidence path. Did the agent identify a meaningful signal? Did it consult the relevant state? Did it explain why the information was actionable? Did it trigger an approved action or merely recommend one? And did it avoid escalating from weak evidence to high-confidence language?^[2]

Start with recommendations and human gates

For a new proactive workflow, the safest first version is usually an agent that prepares a recommendation with evidence and lets a person decide what happens next. This lets the team learn whether the agent notices the right things without giving it authority to create tickets, change code, notify customers, or alter a production system on its own.

1. Choose one narrow signalFor example: a failing release check, a recurring support issue, or a newly introduced dependency risk. Do not begin with “monitor everything.”
2. Define what deserves interruptionWrite both positive and negative examples. Be explicit about when the agent should gather more evidence or remain silent.
3. Require an evidence bundleA recommendation should show the relevant observations, source timestamps, confidence limits, and the smallest reasonable next action.
4. Keep the human decision visibleLet the person accept, decline, defer, or correct the recommendation. Those responses become the feedback needed to improve the insight policy.
5. Measure interruption qualityTrack useful interventions, false alarms, missed opportunities, time to action, and whether users routinely ignore or override the agent.

The best proactive agent is not the most talkative one. It is the one that earns attention: it notices the right change, grounds its reasoning, chooses an appropriate moment, respects the human’s authority, and learns when silence is the better response.^[1]