“We’re living in a multi-agent simulation ourself, and so if you want agents to act in the world and work together with humans, they would also have to work together with each other.”
Multi-agent interaction is essential for advancing AI, enabling specialized agents to collaborate on complex, real-world tasks.
In this conversation, Karthik Narasimhan (Sierra and Princeton) outlines a three-part framework for language-enabled agents: combining reasoning with action, learning new skills through language, and controlling agents via natural language with proper guardrails. He also touches on how building agents for real-world deployment differs vastly from creating demos.
He discusses creating TAU-Bench and SWE-bench as evaluation frameworks for customer support and coding agents respectively.
He also highlights how multi-agent interaction is essential for AGI, the need for continuously learning agents, and how to think about reinforcement learning paradigms.
And finally, Karthik offers his advice for students interested in research, and working in the industry.
Watch the Panel Discussion
Corinne Riley
Corinne works with early-stage founders who are creating data and AI products at the infrastructure and application layers.
visually hidden
