Research
Peer-grade work on how autonomous agents run: the serving layer, the context window and the loop. Published openly and free to read.
One Call, a Whole Job: Testing a Model With Agents Inside
We sent three real jobs to the Sentinel model as single API calls, 9 calls in all, and checked the work ourselves. 9 of 9 passed. Median 3.9 min and 4 tool calls per job, on Huihui-Qwen3.8-27B-abliterated:UD-Q4_K_XL.
Two Sessions at Once: Isolating Parallel Turns in an Agent Dashboard
A second session sent 1.5 seconds after a long one answered in 3.3 seconds while the first was still writing, and not one event crossed between them. What it took: two model slots and a stream that belongs to its session.
Prefix Stability: Why Long Agent Sessions Re-read Themselves
One set of weights, one 8,180-token session, two ways of assembling the prompt. A deterministic prefix is served from cache in 0.5 s; the sliding-window prefix is re-read in 44 s. The model never changed.
Context Is Not Free: Generation Speed Against Window Size on Local Models
The same 27-billion-parameter model generates at 18 tokens per second with 10k tokens of context and 5.3 with 50k. Bigger windows made long sessions slower, not smarter. The measurements, and the window we chose.