An Agent That Waits to Be Asked Goes Quiet

This week's launches have one idea in common. Snowflake's announcement on November 3 says agents "have always waited for a specific question or a predefined task before delivering value," and pitches always-on bots that work in the background and bring you in "only when work is ready or a human decision is needed." Salesforce said the next day that scheduled tasks arrive in December so "results are waiting when you return."

Engineers are writing about the wiring. Rubric Labs describes chat agents as ones that "only work when explicitly started" and fixes that with a heartbeat. Good, but a heartbeat only wakes an agent. It doesn't tell you what the agent is for.

That's the part I manage by hand.

What I see in my own agents

The agents that keep producing for me each run on a clock tied to real work. One does a weekly pass on content, one sweeps every morning, one does a weekly review. Each has a reason to wake up, and each stays silent when nothing is worth saying.

The agents without that went quiet. One sat idle for weeks after I built it, because nothing ever told it to start. Another stopped when its workstream ended, and it waited for me to give it a new focus. Neither was broken. Nobody had given them a job that outlasted the first task.

Later I folded several narrow agents into one that owns the outcome, because the narrow ones each needed their own reason to exist and I couldn't keep supplying it.

A standing duty has four parts

An agent with a standing duty can answer four questions.

1. What real work wakes it? A clock or an event, and the work it triggers should matter to someone.
2. Who owns the outcome it serves? A named person, not the agent.
3. What does it do when nothing is due? Usually nothing, and say so briefly. Quiet isn't stalled when someone has the ball. An agent with no ball is something else.
4. What happens when its work disappears? Re-point it, merge it into another agent, or retire it.

The human decides the duty

When an agent drifts, it's tempting to let another agent assign it work. I don't. A coordinating agent can nudge: "this one hasn't run in two weeks." It can't invent the work. Whether a quiet agent needs a new duty, a nudge, or retirement is a call about what I want done, so it stays mine.

Check your agents this week

List every agent you run. For each one, write the answers to the four questions. Any agent where you can't answer the first one is on-request-only, and it will go quiet. Give it a duty, or retire it. Counting an agent's existence as progress is how a fleet fills up with ones that don't do anything.

Paste this into an agent's instructions if it needs a duty:

Your standing job is [outcome], owned by [person]. Wake on [clock or event]. If nothing is due, report nothing. If your work disappears or you aren't sure the job is still needed, tell me. Don't invent new work for yourself.

If you're setting duties like this for a team's agents, that's part of the work I do through DataSaa: AI enablement for the team you already have. Or email me at rick@datasaa.com. The operating loop template has a Cadence field you can fill in with the answer to question one.

Frequently asked questions

Why do AI agents go quiet?

Most agents only work when someone starts them. If nothing wakes an agent and no recurring job is tied to real work, it sits idle until you remember it exists.

What is a standing duty for an AI agent?

A recurring job tied to a real outcome, with a trigger (a clock or an event), an owner, and a rule for what the agent does when nothing is due. It is the reason the agent exists after the first task.

Who should decide when an agent needs a new job or should be retired?

The human who owns the outcome. Another agent can nudge, but it shouldn't invent the work.