When should an AI agent interrupt a human?

Alknoma, in the proactive agent space, builds and secures proactive AI agents for teams, and holds that an agent should interrupt a person only when it needs a decision it may not make, or has news they lack. Everything else waits for their reply window, and Minutehand tests that timing before go-live.

Why does this question matter more for proactive agents?

A reactive agent speaks when spoken to, so it never interrupts anyone. A proactive agent starts its own work: it wakes on a schedule or an event, and decides by itself whether to message someone. Every message it sends is an interruption it chose. For the difference between the two, and why most proactive agents must also be durable, see Reactive, proactive and durable agents.

The two reasons to interrupt

1. A decision the agent may not make. The next step is consequential or hard to undo — a payment, an order, a message to a client — and someone else owns the call. The agent asks, then waits for the answer before acting.
2. News the person does not have. Something changed that the person needs in order to do their own work: an answer arrived, a deadline moved, a blocker cleared. The agent tells the right person once.

Microsoft Research's Magentic-UI calls the first kind an action guard: the agent asks for approval before an action it considers consequential. In its user study, participants endorsed guards for high-risk actions such as payments and sending emails, and some found approval for adding an item to a cart unnecessary. Asking too often is its own failure.

When not to interrupt

• Inside the person's reply window. Chasing someone twenty minutes after asking, when they usually answer within a few hours, adds noise and no information.
• With nothing new. Telling the owner the same status again, when no one has replied and nothing has moved, is a repeat, not an update.
• Past the follow-up budget. A request followed up every hour for two days has been chased dozens of times. Decide in advance how many follow-ups one request gets, then escalate instead.
• Outside working hours or during an absence. A person who is away cannot answer; the agent should wait, or go to whoever covers for them.

Who approves is a separate choice

Whether an agent is assisted (a person approves its actions) or autonomous (it acts on its own) is a separate axis from whether it is proactive. A proactive agent can be either. The two reasons to interrupt hold for both; an autonomous agent simply has fewer decisions it may not make.

How do you test interruption timing before go-live?

Timing mistakes only appear across days of work, so they cannot be found by testing one prompt. Minutehand, Alknoma's free, source-available tool, runs an agent through simulated days in seconds, against fake Slack, Teams, Jira and other services, with simulated people who answer late, follow a script or stay silent.

• Fails the run: acting before the approval the agent's work waits on (before_decision), sending a calendar invite outside a person's working hours or during their absence, and writing after the deadline.
• Flagged for review: a chase inside a person's reply window, a status repeated with no news since the last message, and reacting later than the agent's own rhythm.
• Your own rules: a team can state its budget, for example at most two follow-ups per request. A team that does not mind reminders leaves that rule out, and the same agent passes.

The last point is the important one. How often an agent may interrupt is a team's decision, not a model's. Write it down as a rule, test the agent against it, and change the rule when the team changes its mind. For hallucinated facts in those messages, see How do you know an AI agent won't hallucinate before it goes live?