Infer-forge runs a 38-part serving project on autonomous agents
- Source
- x.com
- Date
Sharing a new collaboration with the #SGLang community. https://t.co/Ef5HyfGZPF Infer-forge uses Harness + Loop + Graph to keep 28-hour median task lifetimes fully traceable. Peak tasks in flight: 2→9. A full DeepSeek-V4-Pro serving project split into 38 pieces, each verified like a human engineer. And the harness catches kernel silent corruption on its own.
Context
Ant Open Source, working with the SGLang -engine community, described Infer-forge, a system for running long, multi-step engineering projects with autonomous agents. In one project cited, an entire DeepSeek-V4-Pro serving optimization effort was split into 38 separate pieces, each independently verified the way a human engineer's work would be reviewed, rather than accepted on the 's own say-so.
Ant reports a median task lifetime of 28 hours that stayed fully traceable throughout, with the number of tasks running at once scaling from 2 to 9 over the course of the project. The post also says the itself caught 'silent kernel corruption,' a class of bug where GPU kernel code produces a wrong numeric result without crashing, otherwise difficult to detect without a dedicated verification step. Ant did not detail how the three named components, Harness, Loop and Graph, divide that work in this post, so the specific mechanism behind the traceability and detection claims remains unexplained here.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Checking sign-in…
Loading comments…