I stopped counting agent sessions long before ten thousand.

My businesses run on AI agents the way other small businesses run on a couple of part-timers. They watch flight fares, draft newsletters, render ad creative, answer Instagram comments, and file reports while I sleep.

This isn’t my first time delegating the daily work, either. In my content and affiliate marketing days I ran virtual assistants in the Philippines. Good ones. They covered my 3am just fine. The catch was they were asleep during my afternoon.

AI agents are proving just as capable, often more, with less training and a fraction of the cost. And the real advantage: they work all 24 hours, not just the opposite 8.

Ten thousand sessions is enough reps to know what actually breaks.

Spoiler: it’s almost never what people worry about. And every failure in this post is real, pulled straight from my own logs.

Agents fail politely

Nobody warns you about this one.

When a human employee can’t do something, they tell you. When an agent can’t do something, it often says “done” and moves on.

My favorite specimen: a notification function that a past session had replaced with a stub. It returned success. It logged a cheerful message. It sent nothing, for months. The deal watcher ran happily on top of it the whole time, convinced it was keeping me informed.

The day I found it, the fix took four minutes. Finding it took an audit.

The lesson is not “agents lie.” Although they sometimes do. The deeper problem is quieter: agents inherit the system exactly as it is, including the parts you forgot you changed. Every one of those no-op calls was faithfully logged. A human would have said “hey, this notify thing doesn’t seem to do anything.” An agent needs to be told to check.

So now I audit. Not the agents. The assumptions.

The most productive session type in my history is the one where an agent walks my own project tracker claim by claim and tests each “next action” against reality: the live site, the git log, the running process list.

The last audit found a third of my open items were already done. And one “working” system that was silently half-broken. The tracker was fiction in both directions.

The boring failures are the expensive ones

In ten thousand sessions, I have never lost a dollar to an agent being dumb in the way headlines predict.

No rogue purchases. No unhinged emails. No deleted databases. The approval gates make those structurally impossible, and I wrote about that pattern separately.

What actually costs money is boring:

  • An image API that caches an error forever unless you add a cache-busting parameter.
  • A tunnel provider that silently stops being accepted by a platform’s image fetcher.
  • A login that expires mid-run, so the fallback path quietly produces a worse newsletter.
  • An agent politely checking an empty inbox on a schedule, burning money to discover nothing, hundreds of times.

Every one of those is a systems problem wearing an AI costume.

The fix is never a smarter model. It’s a dumber component. A zero-cost scheduled job does the watching. A health check compares matched-versus-sent counts. The expensive, smart agent only wakes up when there’s judgment to apply.

Judgment is the only scarce input

Here’s the real shape of the whole operation.

Agents made execution nearly free. Rendering 122 web pages, extracting a year of fare history with every number source-checked, staging ad campaigns in three formats, drafting this post. All cheap now.

What didn’t get cheaper:

  • Deciding which of those things to do.
  • Catching the one extracted fare that was subtly mislabeled.
  • Knowing a deals page should show fewer deals, so the product feels rare instead of common.
  • Saying no to the feature that pattern-matches to progress.

Twenty years of sales taught me that pipeline is a budget for attention. Agents didn’t repeal that law. They raised the stakes on it, because attention is now the only line item left.

The operating model that survived ten thousand sessions is simple: agents propose, humans dispose. Everything ships through an approval queue. My yes is the bottleneck, on purpose.

A bottleneck you choose is called quality control.

The scoreboard

Ten thousand plus sessions. Two newsletters, an alerts product, an ads pipeline, and a game in packaging.

Zero catastrophic agent failures. One human, several robots.

And a growing file called operations.md where every polite failure gets a line, so it only costs me once.

If you’re starting your own fleet, skip ahead: write the gotchas file on day one, make success claims verifiable, and let nothing say “done” that can’t prove it.

Agents don’t fail loudly. They fail politely, in writing, and the audit is where you find out which.