Cerebras Supernova: Silas Alberti (Cognition) on Why Cloud Agents Are Finally Ready

Two years after Devin launched, is this finally the moment for long-running cloud agents? At Cerebras Supernova, Sarah Chieng sits down with Silas Alberti, who leads research at Cognition, to trace Devin’s shift from succeeding on roughly 30% of tasks to about 90%, and why that reliability changes which product form factor makes sense. Silas explains why cloud agents that run for hours in parallel only work once you can leave them unsupervised, how companies well beyond tech are adopting software-engineering agents, and why very fast inference keeps users in flow. Chapters: 00:00 Agent reliability jumps from 30% to 90% 00:18 Why Devin was ahead of its time 02:11 AI agents and quant trading 03:00 Enterprise adoption beyond tech 03:56 How Silas uses AI personally 04:52 How Cerebras unlocks faster agent workflows 05:19 The semi-async valley of death 06:21 Fast inference helps developers stay in flow