AI agents are becoming cron jobs. The hard part is uptime

On Thursday, Nicolas Bustamante, who works on AI for knowledge workers at Microsoft and previously ran Fintool before Microsoft bought it, posted an observation that anyone using AI tools this month has probably felt without naming it. @nicbstme wrote that over the past month most agents and AI apps have started setting reminders, scheduling background tasks and checking back later. His summary: "Basically cron jobs at scale."

A cron job, for the non-engineers, is the oldest trick in computing. You tell a server to run something at a set time, every hour or every Monday at 7 a.m., and it does it whether anyone is watching or not. Bustamante's point is that chatbots now do this too: they keep working on a schedule after you stop talking to them.

Why agents are learning to come back later

He gives two reasons, and they're worth separating.

The first is practical. A lot of real tasks take longer than one conversation. His example is a restaurant reservation: the useful agent doesn't answer "no tables" and stop. It checks every hour and tells you when something opens. Or, working in a spreadsheet, it says it will update the workbook once a company reports earnings.

The second reason is commercial, and he's candid about it. In his words, "AI has a retention problem. People forget to open the app." An agent that comes back with an update you care about gives you a reason to return. So part of this trend is product teams chasing engagement.

For me the most important part is near the end. Bustamante says giving an agent a scheduling tool is conceptually simple: hand it the tool and tell it in the prompt that it can use it. "The hard part is reliably running billions of scheduled jobs at scale."

The follow-up post: agents messaging agents

A few hours later, @nicbstme posted a funnier and more revealing scene from his own work. His agents run in a loop toward a goal. When one needs context, it messages a colleague on Teams. Some of those colleagues run agents too, so their agent answers his, after digging the fact out of emails, documents or meeting notes that neither person remembered.

He was getting coffee when a coworker got a Teams message "from me," looked up, saw Bustamante wasn't typing, and figured it out. His conclusion: "We are already becoming managers of agents talking to other agents." He thinks the next step is a company brain where context lives with access rights attached, so an agent can fetch only what it is allowed to see.

My read

Both posts describe the same shift. More and more, what makes an agent useful is the work it does while you aren't looking, and a clever answer in the moment counts for less.

I agree with Bustamante that scheduling is the easy part and reliability is the hard part, and that holds at small scale too. We run scheduled AI jobs for our own site. When they fail, the cause is rarely a bad answer from the model. It's usually a network drop early in the morning, an expired login, or a job that quietly didn't fire and sat unnoticed. When an agent promises to check back tomorrow, somebody has to know whether it did.

The agent-to-agent story also has a sober side. It works because his company has permissions and an audit trail around those Teams messages. Without that, an agent answering on your behalf from your inbox is a risk you'd have to manage.

What this means for a small business

Most small businesses run on things that need checking back: the quote that went quiet, the invoice that's 10 days late, the customer who said "call me after the holidays." Those are exactly the tasks Bustamante is describing. They're cheap to automate on a schedule and easy to forget by hand.

Before you hand any of them to an agent, answer three questions:

• What should happen, and when?
• Who finds out if it didn't run?
• What is the agent allowed to see and say on your behalf?

If you're not sure where your own follow-ups are leaking, our free process audit walks through your week and points to the check-back tasks worth putting on a schedule, along with the ones that should stay with a person.