MarathonBuild：面向长时运行Agent的自适应推理基础设施

We are introducing @MarathonBuild, incubated by @GoKiteAI: adaptive inference infrastructure for long-running agents.

Intelligence keeps getting cheaper, but the monthly bill does not follow. Agents live in that gap: they run long jobs, retry, and plan across many steps, and most of that work does not need an answer this second.

▷ Pick a completion window per request: now, soon, later, or anytime. The longer a job can wait, the more room the scheduler has, up to ~65% less.

▷ Same models, same API. Live with the top five open-weight models, and a one-line install for Claude Code and Codex.

▷ One pricing rule across every model. No per-model negotiation.

Kite gave agents a verifiable identity and the right to pay. Marathon pushes that one layer down, to the compute itself. An agent that calls and settles on its own also needs the cost of that call to be programmable.

That is the loop closing: agents that call for themselves, and pay for themselves. 🪁