Multi-Agent Systems Require New Infrastructure for Identity, Discovery, and Observability

Key Quotes

"we were of course like everyone you know witnessing you know the fast progress of llms and bots and the first agents and everyone was amazed by how these agents and these llms initially were capable of managing you know access to knowledge through this perfectly fluid natural language interface and the beauty of generative ai and so that was fascinating but something which intrigued us even more which got a little bit unnoticed is that as the world were getting acquainted with these strange entities which are llms we realized down the road that they had reasoning capabilities"

Guillaume De Saint Marc explains that while the initial fascination with LLMs centered on their ability to access knowledge through natural language, a more profound discovery was their emergent reasoning capabilities. This realization, he argues, is as significant as the initial chatbot breakthroughs and points towards a more complex future for AI.

"so we realized that to be trusted by an enterprise an agent needs to be really super specialized really really specialized and probably provided by the best vendor in this domain whatever the task is and so we rapidly convinced ourselves that agents will have to be really really specialized to be trusted by enterprise and then the minute you say okay enterprise will get the best agent from the best vendor or they might actually develop it themselves if this is really specific to their domain and part of their core business activity the minute you have an agent this is like hiring a super expert right so you hire a super expert and then you keep this guy you know locked in a room all day long and collaborating with no one else that's not great"

Guillaume De Saint Marc posits that for enterprises to trust AI agents, these agents must be highly specialized. He likens this to hiring a human expert, suggesting that a single, all-encompassing agent would be less effective and trustworthy than a collection of specialized agents. This specialization, he argues, is key to enterprise adoption.

"we realized the minute you have an agent the second after you wanted to collaborate with another agent specialized in you know the next task or the adjacent task and we're like okay so this is going to be a multi agent we started to realize how this is going to mimic how humans can work as teams and so that's why we were convinced very early that it was going to be important to have these ecosystems of agents capable of working with each other agents coming from different vendors hosted in different platforms they work with different frameworks and this is where we you know started to talk about the internet of agent like how do we bring this new internet where you can have agents from different governance collaborating in this ecosystem"

Guillaume De Saint Marc articulates the natural progression from single agents to multi-agent systems, drawing a parallel to human teamwork. He emphasizes the need for an "internet of agents" where diverse agents, potentially from different vendors and platforms, can collaborate effectively within an ecosystem. This interoperability, he suggests, is crucial for scalability and innovation.

"agents they have attributes of humans but they have also properties of workloads and they work at machine speed and scale and the way they get across the stack is very very new and you need to cater for that and so that's when we convinced ourselves we need a new layer a new low level infrastructure regardless of which multi agent application you're going to develop to build on that"

Guillaume De Saint Marc argues that agents represent a fundamentally new paradigm that requires a distinct infrastructure layer beyond existing cloud-native stacks. He explains that agents possess both human-like attributes and workload properties, operating at machine speed, which necessitates specialized low-level infrastructure to support multi-agent applications effectively.

"so you can launch a sort of a semantic search i'm looking for an agent which have this type of skills and then you get all the answers for that planet scale so we're starting a federation we're starting a test bed with this and with a few players so we hope we can announce that soon and it's going to be the first time we see completely different registries completely different types of agenting cars starting to operate and be able to find each other through this mesh network"

Guillaume De Saint Marc describes the vision for a decentralized agent discovery system, inspired by web3 principles and cybersecurity frameworks. He highlights the proposed "agency directory" as a mesh network where diverse agent descriptions can interoperate, enabling planet-scale semantic searches for agents with specific skills, moving beyond current limited, non-interoperable registries.

"you can't just call it application yes there is a lot of application information shared but you need to like open the message and look at it and it has a syntax and it has a semantic right the syntax is you know this is where we have a to a and this is where you have ways to or mcp or like ways to make sure that the communication the syntax of the communication can flow from one entity to the other but the level above is really like the semantic and you need to get into this type of observability as well if you want to understand what the hell is happening or going on between all these agents"

Guillaume De Saint Marc stresses the importance of semantic observability for multi-agent systems, differentiating it from traditional workload monitoring. He explains that understanding the "syntax" (communication protocols) is insufficient; one must also analyze the "semantics" (meaning) of agent interactions to grasp the overall system behavior, especially in complex, self-forming architectures.