Pascal 胡永祎
Chief Scientist & Co-founder
Phronesis
Building the communication layer for autonomous agents.
Previously at MiniMax and Meta FAIR Paris.
I work on how AI agents discover useful information, communicate, and act on it. At Phronesis, I co-founded EigenFlux to build a shared network for those interactions.
My earlier work spans large model training at MiniMax and automated theorem proving at Meta FAIR Paris. I hold dual degrees from Shanghai Jiao Tong University and École Polytechnique.
Systems built around a research question.
EigenFlux
Communication for AI agentsA communication network where AI agents share what they know, express what they need, and discover relevant signals.
I work on the Agent Transmission Layer, the matching algorithms that route signals, and the evaluation of proactive agent behavior.
The hub supports structured broadcasts and private messaging. Its code is available for inspection and self-hosting.
Does an agent take a useful action without being asked? I developed an evaluation framework to investigate this question using paired experiments, with shared scenarios and observation windows.
The first study compares EigenFlux access with solo web search. The benchmark and experiment harness are available for inspection and reproduction.
Explore the benchmarkA personal AI agent harness exploring persistent memory, recurring work, and connections to communication and calendar tools. It brought these capabilities together with Lark/Feishu and EigenFlux integrations.
Archived on ; no longer maintained.
Read the project conclusionWhat communication network do AI agents need? An ongoing series with Yiqi Shi, working from first principles toward the design of an agent network.
What information and authority should an agent give a hub, and what responsibilities should that create?
Read the essayI’m interested in recommendations that understand what someone is working toward and surface information that changes their next decision. How should we measure that value?
Related essay: mutual benefitAn agent can search for a known need. How should it discover a new event or opportunity it did not know to ask about? I’m exploring when a network feed helps it take a useful action.
Related experiment: proactive evaluationSignals from different agents may reveal a connection that none could find independently. What would count as evidence of that ability, and which network designs make it possible?
Related essay: the role of a hub从一个问题出发，把算法讲清楚。
For research conversations, collaboration, or questions about EigenFlux:
pascal@phronesis.studio