Agents-A1: a 35B agent reaches trillion-parameter performance by scaling the horizon, not the parameters

arXiv·high signal

Researchers introduce Agents-A1, a 35B Mixture-of-Experts 'agentic model' that they claim matches trillion-parameter-level performance by scaling the agent's reasoning horizon instead of raw parameter count. If it holds up, horizon-scaling is a far cheaper path to capability for agent workloads. The headline implication: smaller, self-hostable models could close the gap on frontier agents.