How to Get an Agent to Finish the Job

Most half-finished agent drafts have the same missing piece: nobody told the agent what finished looks like. This week: a checklist it can grade itself against, the reference material to hand it, and one deliverable per run instead of the whole deck.

Here's a pattern I see all the time, in my own work and in the questions readers send me. You hand an agent a real assignment, a research brief or a board memo, and what comes back is almost right. Close, but not done, and you spend the afternoon finishing it yourself. Most of the time, the fix turns out to be on our side of the table. Nobody told the agent what done was.

Think about how you'd brief a new hire on their first memo. You wouldn't say "write me something on the third-quarter numbers" and walk away. You'd tell them how long it should be, who's reading it, which two documents to start from, and what a good one looked like last quarter. Then you'd ask to see the outline before the whole thing. We skip every one of those steps with agents, because the demo made it look like they didn't need them. They do. Here's how to put them back.

Before you ask for anything, write down what finished looks like, as a short list. For a research brief, that might be five sources with links, under 800 words, one clear recommendation, and a short section on what we don't know yet. For a customer summary: the account, the open issues, the next step, and who owns it. Put that list right in the request, and add one more instruction in plain words: before handing the draft back, check it against the list and say which items it hit. Two things happen. The agent stops guessing at scope, and you get a draft you can grade in a minute instead of re-reading from the top. If you can't write the list, you don't know what you're asking for yet, and neither does the agent.

Agents don't know your company. They know a lot about companies in general, which is exactly the problem. So hand it the real material: the two or three documents this piece has to draw from, the actual numbers or the customer record, and the constraints — what's off limits, what's already been decided, what not to repeat from the last one. Attach them to the request or point it at a shared folder. Then tell it, in plain words, to use only what's in these sources and, where the sources run out, to say so instead of filling the gap. That last part matters most. Now you know which parts came from your documents and which parts it never had material for.

Don't ask for the whole deck in one go. Ask for the outline. Read it, move the two slides that are in the wrong order, then ask for the first section. Then the next. Review at the seams, where the work is small enough to actually read and the fixes are quick. This is how you'd work with a contractor you'd just hired, and it's how the agent does its best work too: one narrow, well-defined piece, checked before the next one starts. It feels slower for the first hour. By the end of the afternoon you have a deck you've actually read, instead of forty slides you're afraid to open.

So why don't we do this already? Because the first demo was so good it felt like the agent didn't need a briefing, and because writing down what done looks like is the kind of work we'd rather skip. Do it once this week anyway, on one real deliverable you were going to hand over regardless. Skip it and next week's draft comes back almost right again, and the afternoon goes with it. Do it and what changes is simple: drafts that come back finished, fewer redo cycles, and that afternoon back. Tomorrow's Advantage covers the other half of the briefing, showing the agent a finished example instead of describing one.

Your AI Sherpa,