From Connected Agents to Collective Intelligence with Guillaume De Saint Marc of Outshift by Cisco

[Music] Welcome everyone to the Emerge AI and Business Podcast. Today's guest is Guillaume the St. Mark VP of Engineering at ArtShift. ArtShift operates as Cisco's internal incubator for emerging technologies focused on reducing risk and delivering innovative results. Guillaume examines why enterprises building multi-agent AI systems encounter coordination failures that connectivity protocols alone cannot resolve. The conversation covers the infrastructure layers required to enable genuine agent collaboration, including shared semantic grounding, persistent memory architecture and fine-grained task authorization. Along with the security and observability decisions, enterprise leaders must build into their agentic systems from day one. Today's episode is sponsored by ArtShift by Cisco. In this episode we cover why enterprise is scaling multi-agent AI need more than connected agents. To go deeper on this topic and learn how consultants are winning business with evidence-based AI ROI and building long-term capabilities instead of chasing short-term gains, download our 3 PDF report 3 keys to thriving in the coming era of automation. At Emerge.com/cok1. That's emeorj.com/cok1 to download your copy. Now the conversation with Guillaume. Guillaume, welcome to the Emerge AI and business podcast. Hey, you're not the many thanks for having me. It's great to be here. It's great to have someone like you because you're the person that actually has to build the infrastructure that makes all these agentic systems work at scale. I think we're going to get into very interesting things today. I want to start with something that I think a lot of leaders in our audience are quietly sitting with right now. We're not speaking about it out loud because we're not sure exactly where the problem is at the moment. I think your organization's own work on the Internet of Agents was premised on solving a connectivity problem. That means getting agents to find each other, communicate and coordinate. But you've watched this play out from the inside and you've been saying publicly that connectivity is actually the first layer and there's something more in the foundation. There's a fundamental problem underneath that firstly. What is that and what does it look like in practice when organizations hit that wall? Yeah, that's exactly as you've said. Let me start with a little story to try to help everyone understand what we are going after and what is the second layer that you talked about. So, picture, a hospital and not a futuristic one from today. The diagnostic team is excellent. The pharmaceutical team is excellent. They avoid bad drug cocktails. The building team knows every insurance called the patient, coordination team, keeps the schedule running. Now put them all in a room for the first time. With no shared language and no institutional memory of what success looks like. No way to resolve this agreement. And the patient is waiting on the other side at the door. That's not a staffing problem. That's a coordination infrastructure problem that you have, right, with this human. And if we move to the agenteic world, this little story is exactly what enterprise are running into. As they try to move from a single agent to systems of agents working together. And so, if you want the first wave of agente AI work. But the second wave is where things start to break. So single agents performing well-scoped tasks are, for instance, creating a marketing campaign or generating a financial report or coding a new module. All these works pretty well and are obviously very useful these days. But enterprise have real wins-year challenges to be on. And the failure starts when you ask multiple agents to collaborate on something. None of them can achieve a loan. And that's where the wheels come off. And not because the agents are bad at their job on the contrary. But because there is no infrastructure for them to actually work together. So, the way to think about it is so far. We've been focusing on solving the wiring. So, making sure these agents can find each other, pass message, exchange data. All this is happening. But what hasn't been solved so far is what happened once they're in the same room. Put the best IT security team, the best cost management team, or the best customer experience team in a crisis together. If they've never shared a frame for what a good outcome looks like, they will be caught in a deadlock, like the security agent will compromise on nothing, because that's his job. The cost agent will say, the fix is too expensive, whatever. The operation agent says, "Wow, we need to roll back now, now, now." And the customer experience agent will say, "Well, you know, customer can't wait. What are you guys doing?" And all four are writing in a way, but none of them can resolve it without something in the room, which can hold all of them accountable for taking a reasonable decision. And that's the gap. And agents actually are hitting this type of problem faster than humans do, because keep in mind that, contrary to humans, they run and they operate at machine speed and scale. So, they have no patience for ambiguity in a way. And so, just to wrap on this, there is really like a few well-known by now patterns into which these agents will fall, again and again, in production when you try to build a multi-agence solution or system. And the first one is, agent can very easily drift semantically and end up not sharing the same reality. They were asked to do one thing and they end up working on the wrong outcome. In a way, they're lost in semantic translation, if you want. Second is they do not converge. So, they try hard, but they do not converge. Completing agent with no shared frame for negotiation can end up in a deadlock, because they all camp on their position and they don't know how to find a good consensus. And the third classic that we see often as well, which can very rapidly happen with a multi-agence system or solution, is they have problematic blind spots or they keep resetting. And this has to do with memory. So, every session starts from scratch, no memory of what was done before, no compounding of knowledge, past experience, no ability for one agent to access, maybe the crucial information that another agent has, but is not able to share. And these are not edge cases. These three typical fail-on mode that I just described, they happen in production much faster than, think when you start to run things at scale. And so, the pattern I keep seeing is, organization get real wins with individual agents, then they try to put those agents in a room together, figuratively speaking, and things start to break. And in ways nobody could predict. And not because the agents are not capable, simply because the infrastructure for them to actually collaborate does not exist yet. And that's exactly the kind of problem my team myself are building to solve. I love the way that you put that, because you start off by painting as a picture of what does look like with the hospital example. And then you gave us the typical symptoms that we could see in our day, because I think we do have a lot of leaders that, they're thinking about it and they're saying, "Yes, what you're saying makes sense, but how do I know if I have that problem?" Because it's something that we're not really talking about yet. So that was a great way to explain it. Painting is that picture. I want to unpack the language thing first, because I read that in one of your white papers as well. Where basically the argument there is that, in the same way that the leap from individual intelligence to collective intelligence in humans was about, it wasn't about individuals getting smarter. It was about the invention of language. And shared context in that language. And the ability to transfer not just information, but meaning, which is, sounding like that is, that is where our problem is. And the argument is that agents are stuck in that pre-language moment, where they can exchange data, but they cannot share the intention behind it. And you've been building to it what we're calling the Internet of Cognition to close that gap. The gap between just connecting versus being in the same room, if we can say it like that. So what does it actually take to move from agents that can communicate to agents that can genuinely collaborate? And what is that demand from the organization's building them? Yeah, that's really crucial. So let me, by the way, unpack this carefully, because that's really important. And for your earlier point about how, you know, language and how the, basically, the human intelligence has been able to almost like scale horizontally and as distributed from intelligence, we need to do exactly the same with agents here. And let me just go back to how the Internet itself got built, because it's also a lot about, in a way, language, protocol, and communication. The Internet didn't scale because we built faster wires. That we have done, as well, since we're in particular, but it's scale because we built smarter layers. And so we have a sort of a smarter layer problem to address here. In the early days of networking, 40 years ago, computers could exchange data, but only if they agreed exactly on how to format every packet, every handshake, every message. It was brutally didn't scale, and, you know, engineers created a shared stack called the OSI stack, open system interconnection, sorry for the dragon. But this OSI stack is very well known in the, the tech industry, and this is, you know, this stack, invented 40 years ago, we're still building on with this stack today. And it has all the seven layers, right, that you'd expect with, you know, from the physical connectivity to the application level connectivity. It has things like Ethernet. TCP/IP, HTTP, all of it. And it was really built for machine exchanging data. And it worked brilliantly. But agents today are breaking this model because they don't just exchange data. When you think about it, they have attributes of human, but they operate at machine speed and scale. So the reason they delegate, they negotiate, they also act on behalf of humans from time to time. And so we've been thinking about how we can extend this OZ stack, OSI stack, with two new layers. And so this is like a big thing. We've been working for 40 years with seven layers. And all of a sudden we have two new layers, right? So layer eight is really about handling agent communication. Typically you'll find things like the A to A protocol or the MCP protocol, which are being discussed quite a lot these days. And so this is about the how agents can be discovered, how they have identity, how they can message to each other or connect in a clean way. But this is just connectivity, as I was saying earlier. It's just the agentic form of it. We call it also sometimes the syntactic layer, right? This is like, do you have the right syntax to be able to even, you know, exchange message? But the layer nine is actually this very, very interesting layer that is giving us a hard time at the moment because we are putting this under control and is the semantic layer. This is the what layer if you want. This is where agents reconcile what things they mean before they act on them and align on the semantic intent and the grounding of their exchange. And when one agent says switch, for instance, and this is a stupid example of it's a very concrete example. When an agent says switch, does it mean a light switch or a Cisco network switch, right? Which are very different things. So they can get lost in translation on things like this if you don't pay attention. And this is not something layer eight or below layer can help with. That's typically a layer nine problem. And it's somewhere agents are making decisions at machine speed. That ambiguity is the difference between a system that works and one that doesn't work. And we have published a great paper on this OSI extension which you can find on archive and it goes much deeper into this. And beyond that, so even with the right communication layers in place, there is a second problem, which is, I touched upon this with the hospital example. Agents don't remember collectively. So every time a group of agents assemble to solve a problem, they start from scratch unless you do something about it. So there is no memory of last week's incident, for instance, no shared understanding of what worked, what didn't work, no compounded knowledge over time. And we call it organizational amnesia for agents. It's one of the most underappreciated limits in production for agentic systems today. And this is really a big deal. And the architecture for solving it is what we call the Co-Connition fabric. It's a policy-governed shared memory layer if you want where context is stored, version created, and made accessible across agent workflow. And the research on this is published. We have a first great application that we've open source actually done by one of our colleague Julia in the team. It's called Mycelium. You can find it on GitHub, Mycelium.io. And Mycelium is great. It's a coordination hub for multi-agent system. And you can connect different agents to it. And it gives them this collaboration room if you want. It assists agents with semantic negotiation, persistent memory, collective intelligence. And that's kind of here and today. This is not something which we are working on. We have released the first version. And it gives a really great idea of how we can very concretely start to address this problem. And then there is the last point which I want to mention. And I talked about the layer nine protocols to share the intent between agents. And I talked about the cognition fabric to have the shared context and avoid the amnesia. And there's a third key element to this architecture and we call them cognition engines. So cognition engines are modules. They are kind of very special form of agents if you want. They live in the fabric. They work for the collective good of the other agents. And they assist with the communication. They accelerate alignment negotiation. They're kind of accelerators if you want. But we also have some of these agents which act as guardrails. And they are more trying to make sure that the Argentine collaboration doesn't go nowhere, doesn't drift or doesn't start doing things which are not wanted. And we've built the first one. It's called CASA, C-A-S-A. CASA is an acronym. It stands for Continuous Agent Sementech Authorization. It doesn't matter what CASA is doing, it's very interesting. It makes sure that agents only access where they are effectively supposed to access for specific tasks that they are doing. Not the session, the task. So it's very granular. Is each micro task handed to an agent or a multi-agent system will make sure that the actions which are being taken are consistent and make sense. For us, let me give you a very quick example. Let's say an Argentic software solution has access to two tools. Currency exchange, you can check the currency exchange rate between two currencies and make international money transfer. If someone says, I want to check the euro dollar exchange rate. And as part of this task, an agent starts to effectively do that. But also trigger a money transfer, you're going to say, why? I just wanted to check the currency exchange. So this is simple, but this will be caught immediately by Casar. And this will be denied. It's like, no, it doesn't make sense, given the task you were just given. And this is, of course, a separate entity which is doing this, not the agent itself, because otherwise you couldn't trust it. So it's only one component, but it demonstrate that the layer we've described, they are not theoretical. We're starting to roll this out. Casar has been released as well and starts to be used in production. And this is just the beginning of a longer roadmap where we keep adding some of these components to enable the internet of cognition. Again, on these, we have papers. We have a lot of reference you can look at. And there is one last thing, which is also really important to mention, which is, what does the-- I mean, you kind of ask the question, what does this demands from organization to embrace all this? And not just the technology, but also the mindset and the considerations. And I would say that there are a few against three, three is a good number. Things always come by three. The first thing is you cannot overlook security, right? So you need to somehow rethink security from the purely agentic perspective. Identity and access management are really important. And what you've built for humans does not translate to agents. Agents are not humans. They act faster, they they get deeper, they operate at the scale, no human workflow ever did. For instance, when a human, when we, you and I, log the morning in our work environment first thing, our device or browser receive a specific token that gives us access to our work environment for 24 hours, sometimes even longer, right? And this is good because we don't want to have to authenticate every two minutes. Sometimes this happens and this is very annoying. I just, you know, sign, you know, I just authenticate, why do I have to log in again, right? For agents that's exactly the contrary, we've made sure that it's a very different approach. We've made sure that they actually keep signing in the system. When we give them a token, it's a one time token. They can't do more than one action. So actually you have agents who are going to have to authenticate and verify who they are several times per second, which a human would be mad doing, right? So it's a completely different pressure on the system and different way to think about identity. So pay really attention to security because this is key, especially for agentic solution where you have multiple agents. The second one is invest on observability because you need it. Agenteic observability is not your typical observability, is not just your classic software observability. You, by all means, keep this one. This is important. But agentic observability precisely gets into this layer mind that I was looking at, it's looking at what the substance of the message and the substance of the, the meaning of the activities that agents are performing. And it's a new discipline. It's very important. And the team building it now will be scaled with confidence. The team that are deferring it and say, I will look at it later. They will struggle because they will have to, they will have to, before they know, they will have to suffer a little bit of consequence. They will have to rearchitect, bring it into running systems, which is an issue. The issues I talked about, you know, the drift, the DEM, Nesia, this type of thing. If you don't have observability, you won't see it. And the last one is a very simple one, quick one, built on open and interoperable foundations. The state of the artist is moving fast for all of us, even for the tech vendors, by the way. So vendor Lockheed is a trap and open foundations is really how you keep the ability to adapt. That's really, really important. I can already imagine the audience thinking, okay, but what if I'm already in that vendor trap? So for an organization that's already deep into a single vendor's orchestration stack, is it still possible for those vendors to alter what you're describing, or have they foreclosed that option completely? - No, it's never too late, right? It's just there is a bit of a cost. It's fine if I could give a few advice for this organization is, you know, whatever situation you're in, try to start with, you know, a real content problem where you're going to try to apply a different approach, which is based on open standards. You know, don't start with a moonshot, because this might be, you're going to have to tackle too many challenges in parallel. But pick a workflow where you can put tools. three agents together on a specific business challenge and look at how they coordinate. And there's good starting point. Typically, IT operations or customer service workflow are good starting point because the failure mode are very visible and the upside of this system working is also very measurable. So again, the goal is not to solve everything at once, but start like this. And progressively, you'll be able to migrate from your logged in solution to something more open. That's a great thing. It's great to know that the option is still there because I think we're going to have a lot of aha moments listening to this. And the problem with the conversation like this is it stays in the conceptual and it doesn't always land in the room where the decisions are being made. And you've been very clear in what you've been describing today that the organizations that will be positioned for the next phase of a gender care are the ones that make certain infrastructure and governance decisions now, not after they have scaled because I think that's when it becomes a bit more tricky to get it back to what we're talking about today. So it's going to require a bit of a mind shift for a senior leader who has agents running in production, real workflows, depending on these agents. And now needs to make an architecture choice that will compound over the next two to three years. What is the most important shift in how they need to be thinking about this? Yeah, so I'm going to point out some of the stuff I already touched upon, but this is a great question because these are really practical advices of things to keep, especially as a decision maker, things you need to keep really top of mind. And these are decisions which your team will need you to make. So in terms of so again, even for those who are starting their multi-agent journey, I'm not just talking about those who might have already cornered themselves a little bit with a proprietary solution. So again, start with one real use case which is important for you, which you understand well, where you can really measure the outcome and see what works, what doesn't work. Again, IT operation, customer service, customer experience, these are great starting points. The second thing is treat security, observability and interoperability as the one requirements. Not day two. We'll add it later is really a bad idea here because the organization that build access controls, observability, and anticipate the need to really operate across different departments or organization with their agenteic systems. From the start are the ones that are able to scale with confidence. The interoperability is key and it connects of course to open and interoperable standards. It's very interesting to see how within a given organization you have two business units or two departments who are already stuck to each other and they cannot even connect these agents or start to make them work together. Let aside connecting with agents from your partners or from your main SaaS vendor, bring your HR agent same with your finance and with your customer or sales management system. So this is really key. Agents are moving faster than a human workflow and so their failure mode will compound faster too. Really recommend to build these capabilities before you realize you need them. The last one, let me say it again, build on open and interoperable foundations. There is a lot of activities in the industry at the moment. So instead of the artist changing fast, if you build on proper foundations, yes, you'll be cornered. So you can't do that. Open foundations are how you really preserve the ability to adapt. Agency that we've funded with so it was a Cisco, Dell, Google Cloud, Oracle and Red Hat initiative within the Linux foundation is a good example. It's a completely vendor neutral. It's a price-grade type of infrastructure for multi-agent systems. So it's a good example of the kind of, I would say, open and interoperable foundation you can consider. AIF, which is the Agents TKI foundation, which is also under the Linux foundation, is another example where we are actively working, not just as our chief and Cisco, but a lot of big AI tech players. So the goal is not to own the internet of agent, right? No single player in own it, right? The goal is to build the infrastructure that makes it work for everyone. And this is important because just like the internet has perwared a lot of the digital economy, we know that such system maximizes the value for the entire ecosystem versus closed ones. And that's really important. To be honest, I know that at this point a lot of what people call multi-agent AI today is still very early, right? Because before you go multi-agent you need to understand how you can go monoagent. And this is just the beginning. But some of it is working, and the industry is definitely heading in that direction to solve really, really valuable business cases. So the Casar module I mentioned is open source on GitHub. It's there. It's here today. You can try. You can use it. My serum is also open source. The research is published for all of these things. If you want to see how the architecture problems we've been talking about actually gets solved, not in theory, but in code, I would say those are the place to start. And very concretely, if you want to to think through the what it means for your organization, well, our chief team is working through exactly these questions. And I would say do not hesitate to reach out because these conversations are really worth having. You can reach out directly. You can join some of them open source work group where we are active, not just as many players. These are all great venues to discuss about the specific use case and how to embrace in your organization. I feel like this is a conversation that I could go on for four more hours. There's a lot to still unpack. And I think we're giving the audience a lot of information. And all of it is super important. But if the audience remembers nothing else, I would want them to walk away with just these three ideas because I think this is where we're really getting that that mental understanding and mental breakthrough of what's happening is first that connected is not the same as collaborative. Agents can talk to each other, but it does not necessarily mean that they're thinking together. So that's definitely the first thing that we need to realize. And then the infrastructure decisions made now will compound open standards, observability and defined agency limits aren't optional extras. And that's what you just said. In fact, these are not afterthoughts. This is actually the foundation that determines whether the next phase of deployment scales or stalls. And I think we're all done with the whole stuck in pilot mode and stalling mode. So that's definitely something to take into consideration. And then the shift is from managing individual agents to governing intelligent workflows. And that requires a new mental model. And it's not just new tooling. It's a whole new way of thinking about it. Again, this has genuinely been a very useful conversation. I think it's the kind of conversation where our audience walks away with you might have ruined someone's entire day, but in a good way, because they know, I'm not able to think about anything else for the rest of the day. It's not just more information. It's more practical thinking. So thank you for sharing how you're thinking about this from the inside. Absolutely. It was a great discussion. Very happy. I've had this discussion with you, Yolande. Rapping out today's episode, I think the three key takeaways from our conversation with you. First, connected agents are not the same as collaborative agents. Interprises must invest in semantic grounding, persistent memory and coordination layers to prevent the drift, deadlock and amnesia that emerge at scale in multi-agent production systems. Second, security, observability and interoperability must be treated as day one architectural requirements. Not deferred enhancements, because agentic failure modes compound faster than those in human driven systems and become significantly harder to remediate once deployments are running. And finally, organizations already committed to proprietary orchestration stacks should identify a single well-scoped use case built on open interoperable foundations, user-to-demonstrate coordination infrastructure principles and expand from there rather than attempting a full scale migration at once. In this episode, we covered why enterprises getting multi-agent AI need more than connected agents. To go deeper on this topic and to learn how consultants are winning a business with evidence-based AI ROI and building long-term capabilities instead of chasing short-term gains, download our free PDF report, three keys to thriving in the coming era of automation, at emerge.com/cok1. That's emearj.com/cok1 to download your copy. On behalf of the team at Emerge, we'll see you on the next episode.