Blog
Notes from
From the watch floor
Notes from
the watch floor.
Essays on operating AI at scale — drift, escalation design, and the unglamorous work of keeping agents honest.
Essays on operating AI at scale — drift, escalation design, and the unglamorous work of keeping agents honest.
A 9B model that generates its own training tasks and builds its own harness — MIT-licensed, local-first, and our go-to sovereign small model over Gemma 4, Qwen, and Nemotron. The recipe, the numbers, and the caveats.
Read the post →The AI frontier split in two — proprietary APIs and downloadable open weights. The Watch now tracks them separately: here is what each side looks like in October 2026, with the receipts.
Read the post →Deploying agents is easy. Operating thousands of them reliably is the differentiator — and the watching has to happen from outside the execution path.
Read the post →Hard failures get paged. Slow degradation gets shrugged at — until it becomes the incident. Why you should alert on the slope, not the threshold.
Read the post →A human who gets a bare alert and a dashboard is not oversight. Full-context handoffs, severity routing, and feedback loops are what make the loop real.
Read the post →See the command center running a live simulated fleet — or join the waitlist for early access.