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

lateinteraction · x · 2026-08-06

The author argues that in current AI systems, including typical coding agents, almost all intelligence resides inside the neural network, even though much of the capability relies on external tools and harnesses.

However, for Reinforcement Learning Models (RLMs) and beyond, the dynamic might shift. The author suggests that increasingly, intelligence will begin to live within the inductive biases of the external harness. Eventually, this learned system could make the harness just as fundamental to the architecture as activation functions inside a DNN.

Original post →

More from coding & agent

coding & agent channel →