The AI Agent Congress opened with a discussion of the terminology the industry is using to describe AI agents today, and how that terminology might be sharpened. Delegates shared the sense that in the future agents are destined to work together in multi-agent systems; we need specialized terms for the different levels of agent that will work in concert. Some agents will be focused on task completion, others will handle task routing, and still others may offer “meta-cognition” – the ability to think about one’s own thinking to improve a system. OpenAI’s experimental multi-agent framework Swarm proposes a “triage agent” that can execute both routines and handoffs. Perhaps theirs will become the leading framework, perhaps not.
Next, delegates debated the topology of agents that we expect to emerge as the ecosystem grows. Some delegates believed we could come to relate to agents as we do to humans – understanding their skills and personalities, and directing tasks according to their expertise. Others imagined that agent capabilities are likely to consolidate with time, so that ultimately every person will have one personal AI agent and one professional AI agent. Or perhaps agents will consolidate even further, similar to the consolidation we’ve seen on the internet, where many searches begin with Google, whether personal or professional. Any of these futures could come to pass, and the one that does will depend on who owns the stack and distribution endpoints. The kinds of stories we tell about AI agents, particularly in movies, will also influence what gets built and what paradigms emerge.
We also discussed standards, protocols and APIs for agents. The diversity of ideas and approaches to agents that we see today is good to start, but standardization will be important in the medium to long-term. Consumers and companies alike will benefit from agent interoperability and from portability of preferences – both of which depend on standards. One reason why we haven’t yet seen standardization for agents take off is that today’s agents aren’t good at calling deterministic functions. APIs aren’t yet LLM-friendly, but need to be for standards to add value.
The final hour of AI Agent Congress unfolded in an unconference format, where delegates tackled topics like agent identity, pricing models for agent products, UI/UX for agents, agent memory, and evaluation & observability.
There’s so much more to say and explore when it comes to AI agents that represent consumers – if you’re interested in these topics and have ideas about how CR should be showing up in the AI agent landscape, we’re eager to hear from you! Say hi at innovationlab@cr.consumer.org.
