An Hour Gone: When Your AI Agent Quits While You're at Lunch

I stepped away for lunch. Twenty minutes, maybe thirty.

I came back to a build that had been pending for 29 minutes. My agent had kicked off a deploy, watched it for a bit, polled the status every few minutes, and then, somewhere along the way, quietly gave up. No retry. No nudge. No "hey, this looks stuck." It just stopped doing anything and waited for a human who was not at the keyboard.

That is at least an hour gone. The build probably finished on its own minutes after I left, and the next three steps that depended on it never happened. I got back, stared at a stalled pipeline, and re-ran the thing myself like it was 2015.

Here is the part that actually makes me angry: time is the one resource you never get back. Money you can earn again. A wasted hour at lunch is just gone. And I guarantee I am not the only one this happens to.

Why agents quit on you

Today's coding agents are built around a person watching the screen. That assumption is invisible until you walk away, and then it breaks in the dumbest possible way:

• They start a long job (a CI build, a deploy, a migration) and then poll it in a loop, burning their own turn doing nothing useful.
• After a few polls they hit some internal limit, decide they have waited "long enough," and give up instead of escalating.
• The work did not fail. It did not error. It just stalled, silently, because the only thing it knew how to do was wait for you.

A watchdog that only catches crashed agents misses this entirely. The agent looks alive. It is alive. It is alive and doing nothing, which is somehow worse.

What should have happened

I should have come back to one of two things: the deploy finished and the next steps ran, or a single message saying "the build has been pending 20 minutes, want me to retry or cancel?"

That is the whole idea behind a supervisor. The supervisor is permanent. The agent is a disposable worker. When a worker hits a long external job, it should hand the waiting off to something that does not get bored, does not burn a turn, and does not give up:

1. The supervisor owns the wait. It watches the build on a sane schedule instead of making an agent sit and poll.
2. When the build finishes, it wakes the next task with the result, so the work continues without me.
3. When the build fails, it requeues with the error instead of swallowing it.
4. When the build is stuck past a threshold, it does not wait forever in silence. It pulls me in, once, with a real decision: retry, cancel, or keep waiting.

That is the difference between "I lost an hour" and "I came back to finished work and one quick question."

Time is the product

HawkeRun exists for exactly this. The point of the tool is not to be clever. It is to give you your hours back. To keep your agents moving while you are away, to kill what hangs and requeue it, and to interrupt you only when a human actually needs to decide something.

A pending build that eats your afternoon is not a quirky edge case. It is the daily tax of running agents that were never designed to work without you standing over them. We are building HawkeRun so that tax goes to zero.

If you have ever come back from lunch to a stalled pipeline and a wasted hour, this tool is for you. Your time should be spent making things, not re-running a deploy a robot abandoned while you ate a sandwich.