New review paper argues code is how AI agents think and act, not just what they produce

A new review from UIUC, Meta, and Stanford argues that code and the surrounding software harness are the real operational core of AI agents.

The paper divides agent systems into model capabilities, infrastructure, and self-generated code, with execution, memory, tool use, verification, and multi-agent workflows as key pieces.

It reframes agent performance as a software-systems problem, not just a model-output problem.

Products like Claude Code, OpenAI Codex, and GitHub Copilot already reflect this code-based agent architecture.