Agent Architecture: Does Intelligence Reside in the Model or the Harness?

AndrewLampinen · x · 2026-08-07

Discusses the distribution of intelligence in current AI agents (like coding assistants). Although much of the system's capability relies on external tools (like Python or API calls), almost all of the intelligence still resides within the neural network.

With more advanced architectures like Reinforcement Learning Models (RLMs), this balance might shift. More and more intelligence could begin to live around the inductive biases of the harness rather than just inside the model itself.

Related event: AI Agent Intelligence Resides in Models, Not External Tools(2 posts)→

Original post →

More from coding & agent

coding & agent channel →