Why code is the actual brain behind AI agents

a new review paper from meta, stanford, and uiuc argues that code serves as the primary mechanism for ai agent cognition, replacing natural language as the bridge between models and their environments. the authors identify the 'harness'—the infrastructure of tools, sandboxes, and execution loops—as the critical component that enables agents to plan, execute, and verify tasks over long durations. this framework categorizes agent systems into three levels: individual code reasoning, reliability infrastructure like program-of-thoughts, and multi-agent collaboration where shared codebases act as a workspace. commercial systems like anthropic's claude code and openai's codex already utilize this structure to automate complex workflows. the paper concludes that as agents rely more on self-generated code, transparent evaluation methods become essential because standard software tests can obscure emergent risks.