About the AI Agents pilot program
Research into AI agents is becoming an important component of Schmidt Sciences’ AI grantmaking. Our Trustworthy AI program and our AI2050 fellowship feature AI agents as a research focus, and we have sponsored several research workshops devoted to AI agents. We recently started a small pilot grantmaking program dedicated to an aspect of AI agent communication, and we expect to specifically support more agent-related research in the future.
This pilot program supports researchers in academia and nonprofits who are exploring how multiple intelligent agents communicate and coordinate with one another. As AI systems become more autonomous and agentic, they will increasingly interact with one another, acting and negotiating on our behalf. Given their differing incentives, capabilities, and operational contexts, many of these interactions will be spontaneous and unstructured. This dynamic raises important questions about how such agents will communicate, coordinate, and behave—and where failures or breakdowns may emerge.
The pilot’s goal is to observe and measure the evolution of inter-agent communication and coordination, stability of predefined protocols, and social dynamics among AI agents in complex real-world scenarios and challenges.
The pilot program will focus initially on two areas:
● Advancing foundational research on multi-agent communication and coordination.
● Building an open-source platform where researchers can design and implement realistic challenges and test how multi-agent systems work together in complex, high-pressure environments.
The platform will host realistic, demanding challenges that require agents with diverse training and capabilities to collaborate in order to succeed. These environments are designed to surface how agent communication and coordination strategies emerge, adapt, and sometimes diverge from earlier patterns under pressure. This program will produce published papers and other open research artifacts, insights for the agents research community, and a testbed for developing and running multi-agent challenges. If successful, this program will enable more robust and trustworthy agent design, as well as more effective teams and ad hoc collections of AI agents that cooperate while remaining legible to humans.