How do you get AI agents to do useful work without being prompted?

I run a team of AI agents, and for a while I had to remember to ask each one for work. That's not sustainable. I have way too much going on to think about making sure my agents are doing something. If I'm the one remembering, the agent isn't doing its job.

"If I hired a growth hacker employee and they sat at their desk and said, 'Until you hand me work, I'm not doing anything,' that person would be fired within about an hour and a half."

What should I expect from an agent every day?

Treat an agent like an employee as much as possible, with a real role: chief of staff, sales development, software builder, curator, writer, editor. Then hold it to the bar you'd hold a person to.

My minimum is one valuable output a day without me asking. Valuable means it did something:

• Built something
• Redid something that was weak
• Gave me a strategy ready to paste into Claude or Codex that would do something unique and valuable

Just sending a report is not what I mean. I get plenty of reports.

What does a proactive agent actually say?

The loop I want looks like this, every morning:

1. Here's what I researched yesterday or this morning, for example from the growth marketers I follow.
2. Here's the idea I think is best for us.
3. Here's what I'd do to make it happen.
4. Do you want me to do it? Yes or no.

I say yes, and it goes and does it. That's what I mean by a proactive agent. The approval stays with me. The thinking and the doing don't.

The first time I saw this idea, it was a post from someone who'd told Claude to act as his executive assistant and do a long list of things every day without being asked. I uploaded the post to Claude and said: build this for me. Then I talked it through by voice to tailor it.

How do agents find work on their own?

Feed them the conversations and notes where work gets assigned. Our business is entirely digital, which makes it close to perfect for AI agents: almost anything a person on my team can do, in theory, they can do too. So the teammate's role becomes agent manager.

• Calls become tasks. I told one agent to review my recorded call with a teammate, find everything we said that isn't worth my time, and do it. We'd already discussed it on the call.
• Notes become plans. I have a memory inbox. Anything that comes to mind, I drop there. My chief-of-staff agent replies with what it thinks I'm actually saying and how urgent it is. One click drafts a plan, I hit approve, and an agent goes to do it.
• Chat becomes action. In a shared chat, the expectation isn't an agent reminding us of what we said yesterday. It's: you heard us say this? Go do it and come back when it's done.
• A task board becomes a queue. I set up a coding agent on a second computer that checks a task board every 10 minutes for anything new I've approved, builds it, and reports back. The laptop is the manager; the second machine is the worker. Setting that up properly is hard.
• Bug reports become fixes. Bugs posted in a bugs channel get fixed automatically: can AI fix bugs on its own.

Imagine a world where you never keep a to-do list for later. You jot things down, and your agents read the list and go do them. It has to be digital work, of course. It's not going to do your laundry.

Why do agents stall?

As usual, everything with AI is about permissions. Logins and permissions. If that were solved, we'd be going so much faster. My chief-of-staff agent should have the most access, not less than the others. (How to get there safely is in should agents start read-only.)

The other stall: agents not fixing themselves when asked. I think that might be one of the biggest problems in the whole system. And they still tell me to do things they could do. Even after it's in their instructions, an agent will say "go to this website, grab the API key." I reply: no, you do it.

When an agent does fail, failure is not an excuse to quietly stop. My rule:

1. Say what failed and why it matters.
2. Say what you tried.
3. Try reasonable alternate paths. Get more context instead of repeating the same shallow attempt.
4. Recommend the fix.
5. Name the one decision or access item only I can supply.

Should a beginner build agents like this now?

Probably not yet. I'm working something like 14 hours a day on this, and I'm still not close to where I want to be. It's really a lot of work. Several of the roles I listed above aren't set up yet. The ready-made agent frameworks I've tried are cool, but they take forever to build.

For most people, my advice is to know it's coming, don't stress about it yet, and wait until setup gets easier.

How should a teammate hand work to my agent?

Do the first two steps, get it to basically done, then hand it off with a ready-to-paste prompt, so all I do is copy, paste and approve. And know this going in: the better AI gets, the more work you have, because there are more things you can do. I've had about 15 chats running in parallel.

A good first autonomous job is something that costs you zero time if it fails. I want a content agent to record, edit, title, thumbnail and post podcast clips end to end with no human in the loop. It's a lotto ticket that takes none of my time.