Agents Over Bubbles

Ben Thompson argues that AI is not in a bubble, primarily because of the emergence of functional AI agents. He traces three LLM inflection points: ChatGPT (2022), OpenAI's o1 reasoning model (2024), and agentic systems like Claude Code and Codex (late 2025). Agents address most of ChatGPT's original flaws by verifying their own outputs without human involvement. Crucially, agents reduce the number of people needed to drive massive compute demand, since one person can direct multiple agents. On the enterprise side, companies are economically incentivized to replace human organizational overhead with agents, fueling sustained capex investment. Thompson also argues that the model-plus-harness integration required for effective agents undermines the commodity-model thesis, positioning Anthropic and OpenAI as durable value-chain integrators. Microsoft's Copilot Cowork, built on Anthropic's Claude rather than a model-agnostic stack, is cited as evidence that compelling agentic products require tight model-harness integration.