Scheduled AI agents for recurring client briefs

Scheduled AI agents run a recurring client brief on a fixed rhythm, so the first pass is ready when the consultant sits down to it. Setting the schedule takes a minute. Whether the arrangement lasts depends on who reads each run, what they do with it, and whether anyone notices when it stops being useful.

This post assumes the task has already been chosen and written down as a method. It covers what comes next: the rhythm, the review, and the upkeep.

Match the schedule to the decision

Set the schedule by when someone needs the output, not by how often the sources change. A weekly market monitor is worth running weekly only if a person looks at it each week and could act on it. If the client conversation it feeds happens monthly, a monthly run with a longer window is the better choice.

Running more often than the reader can absorb has a cost beyond the wasted runs. People learn to skim, and when they skim they miss real changes.

Name one reviewer for every run

Every scheduled agent needs one named person who expects its output and reads it. A shared inbox or a team channel does not count, because each person assumes another has looked.

The reviewer's task is narrow. They check the claims in the run against the sources shown, decide what matters for the client, and add what the sources do not cover. Nothing from a scheduled run should reach a client until that has happened. Why AI drafts still need review before delivery covers what the check should include.

When the reviewer changes project or leaves, give the agent a new reviewer the same day. Pause any agent that has no reader until it has one.

Keep the brief and the sources fixed between runs

You can only compare a scheduled run with the last one if the brief and the sources are the same. That is the point of scheduling it: you read this month's update against last month's, and any difference between them comes from a change in the world. If the wording of the brief also changed, nobody can tell which caused the difference.

You will still need to change things, for example when a client adds a competitor to watch or a source stops publishing. Make each change deliberately, note the date, and tell the reviewer, so they read the next run with that in mind.

Review the agent itself each quarter

Look at each scheduled agent once a quarter and ask whether it is still worth running. Three questions are enough. Did someone read every run? Did any run lead to a decision or a conversation with the client? Has the brief drifted from what the client now cares about?

If nobody has acted on an agent's runs for a quarter, consider a slower rhythm or stopping it. Agents pile up the way recurring meetings do, and every one that runs unread makes it easier for people to ignore the useful ones.

Pause before you delete

Pause an agent when its engagement goes quiet, and delete it only when the work is finished for good. Client work stops and starts. A retainer is suspended for a quarter, or a monitoring brief is shelved while the client reorganises. Pausing keeps the method, the sources, and the history of earlier runs, so the agent can pick up where it left off.

A cap on what each agent may spend is a useful safeguard for the same reason. A schedule keeps running whether or not anyone is watching it, and a cap puts a ceiling on how much an unattended run can cost.

How this looks in Hebno

In Hebno an agent runs daily, weekly, or monthly, or when someone presses run. It is pinned to a client project and the firm's skills, so its sources stay limited to what that project allows. An agent can be paused without deleting it, each one has a spend cap and pauses before it goes over, and runs are shown by agent so a lead can see which ones are in use. The agents feature page shows the schedule and the run view.

The guide to reusable AI workflows explains where scheduled work fits among the firm's other repeatable methods.