Six times "done" was wrong, and how to make agent work checkable

Andrej Karpathy wrote on 2 October that we will spend a lot more time trying to understand what language models produce. Teams running coding agents are already paying that cost. The expensive part is rarely the model getting the task wrong. It is a report that says "done" with the same confidence whether the work is finished or not.

What it looks like in practice

These happened in our own systems:

• An API returned 201 Created and a media ID; the post went live without its image, in four separate tests.
• A Go test stayed green while the bug it guarded against was present, because the result came from cache.
• An alerting pipeline paged every minute or two for one outage, because its dedup key included a timestamp.
• A CI test failed at exactly 10.00 seconds, the client timeout; the runner's intrusion-prevention tool had banned its own loopback address, 127.0.0.1.

None of these was a model failure. Each was a missing verification step. The full write-up, with the habit that catches each one, is on Medium: Six times "done" was wrong.

The habits, in one line each

1. Ask for evidence of the outcome, not the response code.
2. Ask what was not checked.
3. Break every check once on purpose before trusting it.
4. Have a reader who did not write it check claims against sources.
5. When a failure matches a configured number exactly, measure the machine first.
6. Ask for plain, short sentences: vague claims have nowhere to hide.

Where a decision model fits

Many of the checks above are small, typed questions: did this post render the image, is this the same incident as that one, is this alert a repeat, which tool should handle this. A frontier model is overkill for them. A small decision model answers them fast, on your own hardware, with a probability you can set a threshold on.

deemwar is the OSS partner for OpenJevX, a small open decision model: 598 MB, 178 ms median on a busy laptop CPU, HTTP included.

How we help teams

• A decision-model pilot on one decision. Pick one decision your team or your agents make many times a day (triage, routing, dedupe, "is this done"). We run a small decision model on your own data and report accuracy and latency measured on your cases, including the cases where it loses.
• AI engineering pod. We work in your repo and tickets, set up these verification habits in your agent workflow, and hand them over.
• Migration audit. Before you move a pipeline to agents or to a new model, we replay real work through both and compare decisions, not averages.

Talk to us: io@deemwar.com. Tell us one decision your agents get wrong, and we will tell you whether a pilot can fix it.