Introduction
On Wednesday July 17th, 2024, the inaugural event of the AI Agents London meetup gathered at UCL in Euston for an evening discussing the emerging field of AI agents. The event featured three presentations:
- An introduction to the AI Agents community and various interpretations of AI agents
- The announcement of the $1M AI Agents Challenge by Agentplex founders (the host of the event)
- A deep dive into the Camel multi-agent framework by Oxford star researcher turned startup founder Guohao Li
The highlight of the evening was Dr. Li's presentation, which explored AI societies (drawing from Minsky's work). Li touched on the open challenges in developing cooperative role-playing agents (from endless loops to undistorted benchmarking), and potential applications. It also emerged in Q&A that separate research groups appear to be converging on similar concepts, suggesting that agents are ripe for innovation and perhaps that their time has come, rather than being just a passing trend. The diversity of approaches could lead to rapid progress through cross-pollination, and Li's open source work obviously seeks to fuel this.
The theme of Li's talk was 'scaling laws for agents' but no such law was presented, it was given as more of a provocation: how do we scale multi-agent systems? Is it a matter of data, of environment, or something else yet unidentified? If something does lie awaiting discovery, we could see this approach advance in leaps and bounds.
Agent-driven data generation was another topic of interest, and a Q&A touched on a 'meta-agent' ("critic" or "manager") that oversee and optimise the others' performances.
These notes capture key points from each presentation, offering a glimpse into the current AI agent landscape.