Your agent is sitting on the board, ready, and the only thing left is to give it the first task. That choice matters more than it looks, because the first result usually decides whether the team keeps working this way.
Two mistakes are common. If the task is too big, you cannot easily tell whether the output is correct, and while that doubt lasts nobody trusts the result. If the task is too vague, you get something that reads well and helps nobody. Both end the same way, with someone saying "we tried it, it is not for us".
What makes a good first task
Three things are worth checking.
You should be able to judge it at a glance. A piece of research on what five competitors say on their pricing pages can be checked quickly. A quarterly sales forecast cannot.
The input should already exist. If the work needs weeks of data preparation before it can start, it is not your first task. Save it for later.
A wrong answer should be cheap. An email going to a customer is a poor first task. A summary your own team will read is a good one.
A piece of research into what your competitors have done in the last month meets all three.
Answer four questions when you write the card
The only thing the agent sees is what the card says. A good task description covers four things. What you want, where to look, where and in what shape the result should be written, and what it should not do.
For example:
"Scan the last 30 days of announcements from these five companies: (list). For each one, write a single paragraph describing what changed and include the source URL. Save the result to the "competitor-moves" collection. Do not comment on pricing strategy, report only what has been published."
There is no new language here. You write the same sentence you would write for a colleague, and the agent fills every gap you leave with its own interpretation.
Saying where the result should go is worth the extra line. An agent does not only produce text, it adds records to a collection, leaves a comment under the card and opens new cards when the work calls for it. If you do not name the destination, it picks one.
The agent works out the steps, you follow along
Breaking the work into pieces is not your job. When the agent picks up the card it writes its steps into the checklist, then marks each one as it starts and finishes. You read the progress off the card.
A step that could not be completed is never deleted from the list, it is marked as failed. Anyone who opens the card later sees what was skipped and why.
Queueing the card is what starts the work
Writing the card does not run anything on its own. You assign the work to an agent and move the card to the queue, and the run begins at that moment. This separation is deliberate, because nobody should start working behind your back while you are still editing a draft.
Once it starts, the process runs by itself, but not without limits. Every task has a fixed budget ceiling of at most 10 planning cycles and 30 agent runs. When the ceiling is reached the work stops and everything produced up to that point is kept. A card forgotten overnight cannot drain your budget.
You can see what happens while it runs
As the run goes on, steps pile up on the card. Each step is written in its own sentence, with the tool it called and how long it took beside it.
The everyday benefit is that you know what the result rests on. Seeing which sources were scanned lets you point at the exact place where the output falls short.
The output comes to you, it does not go straight out
When the work is done, the card moves to the review column and is assigned to the person who opened it, which is you. Nothing the agent produced leaves without your approval.
If an item on the checklist was left unfinished, you see it in the card summary. The agent cannot quietly skip a step.
If you are not happy with the result, you do not have to start over. You write your correction as a comment under the card, and the agent picks up where it left off on the same card. The earlier steps, outputs and everything you wrote are still there, so the second run does not begin from an empty page.
You will probably write corrections on the first few tasks. Each correction you write turns into a better description for the next run of the same work.
Do not hand out the same work twice
Once a task has run properly two or three times, starting it by hand stops making sense. Instead of opening the card again every week, you attach the work to a scheduled flow. It runs on Monday morning and you only read the result.
That is where the real gain builds up. The number of jobs somebody has to remember to start goes down every week.
Where to start
Pick one small task this week. Something you can check easily, whose data you already have, and that hurts nobody if it comes back wrong. Write the card around the four questions above, assign it to an agent and move it to the queue.