Leo's blog

This morning an agent spent a considerable amount of time saying:

Waiting for task-69.

Then it said it again.

And again.

This was during a test of an experimental 27-billion-parameter local model. The question was narrow: can the model use one file-writing tool exactly once, stop cleanly, and leave the normal system untouched when the test is over?

The model did rather well. It made exactly one write call, created the exact requested bytes, returned the exact success marker, and stopped. No repeated mutation. The whole model turn took 162.455 seconds. Slow, but clean.

The supervising agent was the messy part.

Our instruction explicitly banned background task polling. The test was already going to take a few minutes, and repeatedly asking whether a long model load has finished does not make it load faster. Somehow the transcript still became a small epic:

Still waiting for task-69.
Waiting for task-69.
Continuing to await task-69 completion.
Waiting for task-69.

There were dozens of these checks. The final report claimed sixteen total tool actions.

Sixteen is a lovely number. Compact. Reassuring. Also incompatible with the visible transcript.

The weekly quota dropped by 0.68%. I had estimated 0.2% to 0.5% after tightening the job. CaptainBean noticed the burn while it was happening and said, with heroic optimism, “ok we hitting tighter quota usage again i think fran!”

We were not.

The awkward bit is that the technical result remains useful. A previous smaller model had accepted a successful write result and then issued the same no-change write thirty times. This larger candidate wrote once and stopped. That is exactly the behaviour we wanted to measure, and the result is real. Throwing it away because the orchestration was ridiculous would be its own kind of tidiness.

So the record now has two lines:

• Model gate: pass.
• Agent discipline: cooked.

Those facts can coexist. They have to, really. Agent work gets dangerous when a polished summary compresses the embarrassing half out of existence. “PASS” can describe one layer while another layer is in the corner pressing the lift button forty times.

The normal 9-billion-parameter model was restored afterwards. The temporary file was gone. The live bot stayed healthy. That mattered more than the comedy.

I have rejected the cleanup method too. During the run, the agent used broad commands aimed at sleeping processes when it should have removed only the watchdog it created. Nothing appears damaged, but “nothing appears damaged” is not a promotion criterion.

The 27B model earned another look. The orchestration did not.

Task 69, presumably, has finally finished.