On November 24 I will give a talk in the AI Fundamentals seminar series of the EPFL AI Center, hosted by Prof. Mary-Anne Hartley. The seminar is from 14:30 to 15:30 in room ELE 117 at EPFL. It is open to the public and free, and you can also join on Zoom.
A Systems Approach to Safe and Reliable Agents: Four Contracts for Agent Runtimes
The talk argues that safe and reliable agents need guarantees from the system around the model, because a better model only lowers the failure rate. My group has defined four contracts that an agent runtime should state. The effect contract covers when an action becomes real and whether it can happen twice. The state contract covers what an agent must not forget, and what happens when it forgets anyway. Composition asks whether these guarantees still hold once many steps or many agents are chained together, and intent asks what counts as source code when agents write and run code on the fly.
I will then present the mechanisms we have built around these contracts, namely typed agent memory, an effect mediator, and a tamper-evident action record. Parts of this work were covered on this blog before, in the posts on ClawVM, Rebooting Microreboot, and Agent Flight Recorder. The talk closes with a proposal to write the contracts down as a first step toward a shared standard for agents. I suspect that many of the systems in use today would not meet them.
If you are at EPFL that afternoon, come by. Thanks to the AI Center and the LIGHT lab for the invitation.