The Ezra Klein Show: How Fast Will A.I. Agents Rip Through the Economy?

Ezra Klein interviews Jack Clark, co-founder and policy head at Anthropic, on the shift from AI chatbots to agents—systems that act independently using tools like code execution or web search. Clark shares examples from Claude Code, Anthropic's coding agent, and discusses its rapid integration into work, economic effects, oversight challenges, and psychological adaptations needed for effective use.

Defining AI Agents and Their Mechanics

AI agents differ from chatbots by operating autonomously after receiving instructions, akin to delegating to a literal-minded colleague. Clark recounts using Claude Code to rebuild a complex species simulation—complete with predators, prey, visualization tools, and dependencies—in 10 minutes, a task that would take a skilled programmer hours or days. This works via multi-agent setups: one agent oversees others, spawning subsystems for subtasks like package installation. Key enabler: recent "reasoning" training, where models learn to solve multi-step problems in simulated environments (e.g., spreadsheets, calculators). They self-correct by spotting errors, building "intuition" for dead ends and resets, moving beyond token prediction to world-modeling that includes self-reference.