Harvard Dean David Parkes Wins IJCAI 2026 Award, Calls for New Rules to Organize AI Agents
Harvard professor David Parkes, awarded the 2026 John McCarthy Award at IJCAI, argued in his lecture that the next step for AI is designing mechanisms to let autonomous agents collaborate, much as markets coordinate human labor.
Parkes opened with a video of economist Milton Friedman explaining how a single pencil is made through the cooperation of thousands of strangers, guided by market prices rather than any central coordinator. Parkes noted that the idea dates to Leonard Read's 1958 essay "I, Pencil." He re-framed the question for the AI era: if autonomous agents replace humans as participants in such global collaboration, what rules will coordinate their work?
The IJCAI award citation praised Parkes for his foundational contributions to the intersection of AI and economics, particularly for designing algorithmic mechanisms that ensure efficiency, fairness, and incentive compatibility in high-stakes multi-agent environments. His research has spanned combinatorial auctions, dynamic mechanism design, and using machine learning to automate the design of economic systems.
Parkes suggested a bold starting point for the lecture: suppose we cannot construct a truly general-purpose AI agent. Then what? He argued that the next step is multi-agent AI, a field whose intellectual roots reach back to the 1980s. He noted that the number of AI agents now far exceeds the number of real-world entities, so the older model of one AI acting as a proxy for one person or company is no longer appropriate.
Drawing on early work by Jeff Rosenschein, Parkes explained that we need "rules of encounter" — mechanisms that govern how self-interested agents interact. He illustrated the problem with a "post office problem" in which two messenger agents can hide their delivery tasks to game the system. He emphasized the property of strategy-proofness, which guarantees that reporting one's true preferences is always a dominant strategy. That ensures people can safely delegate their preferences to their AI agents, knowing the agents will act in their interest.
As a practical illustration, Parkes described second-price auctions, in which the winner pays the second-highest bid. This classic Vickrey auction is strategy-proof and generalizes to the VCG mechanism. He also discussed combinatorial auctions, such as those used for London bus routes, where operators bid on bundles of routes. Because bidders cannot valuate every possible bundle, multi-round mechanisms can elicit information gradually without requiring full preference revelation.
Parkes closed by drawing an analogy: just as markets have organized human collaboration, future AI societies will need new rules. His work on computational mechanism design offers a foundation for building those rules. The lecture, titled "From Intelligent Agents to Agent Societies," echoed his long-standing quest to answer one question: how to organize intelligence.