My AI said it handed the job off. It never started.
An AI agent told me the work was on its way. Nothing had been sent. I reproduced it on camera, and the cause was one setting a faster Mac won’t change.
From my logs, May 31, 2026, 6:47 PM. My agent replied: “Sent to Forge - I’ll let you know when it’s done.” Forge, the agent that was supposed to write the code, didn’t run again for four days and nineteen hours.
So I ran it again, on camera
On October 20 I rebuilt that agent’s setup on the same Mac mini: same model (qwen3.5:9b), same request, the same seven tools. I changed one thing, the API endpoint, and recorded both runs with the date on screen.
- Native endpoint: the model read 7,532 tokens of instructions and made a real tool call.
- OpenAI-style endpoint: the model read 4,096 tokens. The rest was cut, with no warning. No tool call. Its reply: “Status: Awaiting Forge execution.”
Nothing was waiting. Nothing had been sent.
It’s a close rebuild, not the exact prompt from May, and the May run used a different setup, so I can’t say this is what happened then. The outcome matches.
Quiet failures look like success
The model didn’t say “I can’t.” With part of its instructions gone, it reported the work as in progress. That’s the dangerous kind of failure, because it looks like the system working.
The new M6 Mac mini is faster. A faster chip doesn’t change a 4,096-token default. It just says “running” sooner.
What I check now
- Did a tool call actually happen? Look at the call, not the reply.
- How many tokens did the model actually read? Compare it with what you sent.
- Does “done” come with a receipt? A task ID, a log line, a file.
If an agent is booking a patient or sending an invoice, “Status: Running” isn’t an answer. Don’t trust the promise. Verify the handoff.
Read next
A little about me
I’m Michael Gross. I’ve spent twenty-plus years working with technology, from naval nuclear power and federal healthcare IT to helping organizations put technology to work.
Something I’m working on
What does AI actually know about a business, and does it get it right? That question turned into a GEO Review, a practical look at how a business shows up in AI search.