External signal·Stratechery·Mar 16, 2026·Ben Thompson·10 min read
Agents Over Bubbles
“agents actively verify the results without humans needing to be in the loop”
Summary
Ben Thompson distinguishes "answer inference" (the model hands a human an answer) from "agentic inference" (the model does a task autonomously, with agents verifying their own work without a human in the loop). Tracing three LLM inflection points — ChatGPT for token prediction, o1 for reasoning, and Opus 4.5 / Claude Code for the first usable agents — he argues the surge in autonomous, task-completing agents is what drives soaring compute demand and undercuts the AI-bubble thesis, with Anthropic and OpenAI becoming the integration points in the enterprise value chain.
Predictions for the future of work
Predicts agentic inference — software that completes whole tasks rather than answering questions — becomes the dominant mode and the main driver of compute demand, making Anthropic and OpenAI the layer enterprises build on. The future-of-work implication is indirect but pointed: the trajectory is toward work performed autonomously, without a human in the loop.
Originally published by Stratechery · Mar 16, 2026
Read the original at Stratechery