What an agent looks like in my business
Every night, one of my agents reads about two dozen files that other jobs wrote while I slept.
App revenue. Social views per account. Who messaged me. What broke.
It decides what matters and writes me one message. Around 8am it lands on my phone. Wins first.
Then the numbers, then the one lever for the day. Under 140 words, because the early versions sent me essays.
Nobody told it which file to open first or what to leave out. It judged. That’s the agent part.
A fixed sequence that runs the same steps every time is a workflow. Much of my system is exactly that: plain scripts that collect numbers.
The agents sit only where judgment is needed. Want to try a first job on safe, sample files? Get the free AI business starter.
Is an AI agent different from a chatbot?
Yes. A chatbot waits for you. An agent can start on a schedule or a trigger, and it can act: open a file, query an app, draft an email, retry something that failed.
My assistant, Garry, is about 25 scheduled agents on my Mac and a small cloud server.
They all read and write the same notes, so each one can see what the others found.
What goes wrong with AI agents?
More than the demos show. In one two-week stretch this year, my system had six real outages. I, or someone on my team, noticed four.
An AI audit caught one. The agent itself flagged one, weeks late. The monitors I had built for exactly those failures caught none.
The one that still stings: for seven weeks, every agent ran on a different model from the one I had chosen.
The code wrote the model name in the log and never passed it on. Everything looked fine.
So I keep a few rules now. “Scheduled” is not “posted”. “Accepted” is not “delivered”.
A report built on a stale source is worse than a report that says “no fresh data today”.
How much should an AI agent be allowed to do?
Less than you’d think, at first.
The part of my system that clicks and types for me refuses five things, whatever the instruction says.
Payments. Passwords. Deploys. Ad spend. Deleting anything.
Anything that goes out to another person needs a confirmation first.
And one rule about models, because cheap ones are tempting: a cheap model may do the work, but a cheap model never judges what reaches me.
Do you need an AI agent?
Try what I call the boring test. Name one thing you do every week, the same way, that you could write on an index card for a capable stranger.
If you can, an agent can probably take it. If you can’t, an agent will only run your confusion faster.
The agent-shaped project I teach is content distribution.
In my first month testing that system on my own accounts, it published 906 posts that got 811.4K views.
Agents that run for many steps also lose track in a specific way. That’s about their context window.
That system is The Social Machine: one account, a repeatable format, your review step and a weekly learning loop, built live with me.
Technical distinction between workflows and agents: Anthropic: Building effective agents.