Agents Over Bubbles
Updated 2026-04-07
Ben Thompson argues here that AI agents are not simply another hype wave, but a real shift in demand for compute, software, and enterprise adoption.
The Important Shift
The essay makes three paradigms visible:
- ChatGPT made LLMs broadly usable
- Reasoning models made them more reliable
- Agents made them able to act
The decisive point is not only the model but the Harness: the software that steers the model, verifies output, and couples it to tools. That is where differentiation starts to appear.
Why It Matters Economically
Agents do not only increase model usage. They increase the duration and depth of work. A single agent can make repeated model calls, use tools, check outcomes, and iterate. That means compute demand rises not linearly, but structurally.
That also explains why the value is showing up especially in enterprise settings. Productivity there is directly monetizable, and coordination is expensive enough that a capable agent creates real leverage.
The Strongest Thesis
At its core, the essay argues against the simple commodity story. Not every LLM becomes interchangeable raw material if the real product is model plus harness.
Connections
- Anthropic - benefits especially if integration quality matters
- OpenAI - similar dynamic, different product and agent access model
- Claude Code - working example of a strong harness
- Vibe Coding - agents push programming closer to intention steering
- AI Evals - once agents act, evaluation mechanisms matter more