The Invisible Workforce

I opened the dashboard and saw the same thing I’d seen for weeks: a handful of cards, most of them waiting on me. The agents had been running every day, the logs showed completed work, and still nothing moved forward. I figured the autonomy just wasn’t taking hold yet.

Then one of the runs surfaced a different number. Not the count of cards, but the count of runs that had finished and marked themselves “pending review.” One hundred and ten. One hundred and ten pieces of work sitting in a state the surface never showed.

The status wasn’t lying on purpose. It was just reporting what the agent believed had happened. The pipe that was supposed to turn that belief into something I could see had never been wired. So the work existed, the review never did, and I kept thinking the agents weren’t delivering.

That’s the part that sticks. When your agents look idle, the first thing to check isn’t their effort. It’s whether the thing they’re reporting actually reaches you. A status is only as good as the surface that renders it. If the surface is missing, the status becomes a private claim the boss never sees.

I’ve started treating every “done” or “waiting” as a telemetry event instead of a promise. The run either emitted an artifact I can open, or it didn’t. The card either appeared in the queue, or it didn’t. Everything else is a story the agent is telling itself.

The practical move is small. Before you ask why the agents aren’t moving faster, pull the last ten runs and ask what actually left their hands and where it landed. If the answer is “it should have shown up here,” you’ve found the real constraint. Fix the pipe before you rewrite the prompt.

First in the Building with Agents series — real lessons from running a company on AI employees. The rules these incidents produce live in the open at Harness.