David Cushman — Executive Research Leader, HFS Research[00:21]
Welcome to this HFS video cast. Today we’re talking about what’s rapidly becoming the real architecture problem of enterprise AI. Not which model or co-pilot to select, but how to run multi-agent systems reliably, safely, and at scale. And joining me is Babak Hodjat. He’s the Chief AI Officer at Cognizant. Babak’s been right in the middle of building and deploying applied AI in complex enterprise environments, where the hard part is no longer getting a demo to work, it is making systems dependable in production, handling real constraints, messy data, real users, real risk, real accountability. So the conversation today is anchored in a new field guide we’ve just published in association with Cognizant, learning from early leaders in agentic enterprise, a group we’ve called orchestrators. These are organizations that are already running five plus agents in production. And what surprised us is the pace of takeoff. The most mature deployments typically run around a dozen agents, sometimes 20. And I suspect since we did that research, even more. Nearly a quarter of orchestrators are moving from proof of concept to production in under three months. And I suspect, again, changes in the market, such as the arrival of OpenClaw, is compressing that cycle as well. But the paper’s central learning is even sharper. The biggest failure mode isn’t model quality. It’s systems and operating model debt. And once you cross into multi-agent reality you can’t just add agents. Complexity compounds, coordination breaks and post-live surprises show up such as emergent behaviors, cascading failures and auditability gaps. So the thesis is you don’t need another agent, you need an agent operating system. A unified layer for governance and autonomy, orchestration, observability, explainability, data trust, and a human agent workforce model. So, Babak, I want to use this as a practical conversation to dig into what’s real, what’s hype, where you see multi-agent systems actually heading, especially as enterprises try to move from isolated pilots to cross-functional automation at scale. So kicking that off, in the paper, we have orchestrators reporting some real post-live landmines emerging behaviors, cascading failures, and auditability gaps come up about one in five instances. And so from your vantage point, where you’re actually digging deep into this, where do multi-agent systems actually break first in an enterprise production? And what’s that telling us about what progress should mean right now?
Babak Hodjat — Chief AI Officer, Cognizant[03:15]
Yeah, great question. I mean, we’re also experiencing this within Cognizant. We’ve created a multi-agentic system of hundreds of agents within Cognizant that’s actually working quite effectively. And we’ve learned on the job as to what are the failure points as well. And there are multiple. Expecting that coarse-grain semi-autonomous agents would ad hoc be able to coordinate with one another and resolve issues is probably naive. In fact, one way to reduce some of the issues that you just mentioned is to actually be more granular and more sort of tailored and engineering these systems and the orchestration between the agents themselves. So that’s one path that is there. The other that you also alluded to is the safety and trust layer, which demands some form of centralized control. While it is very important for us to design for empowerment so that various different elements within our business can bring in agents, even define new agents and plug them in, there has to be a process for that, for registering the agents, for safeguarding them, to actually applying guardrails, to applying policies, and for measuring trust in the system as a whole and in its subcomponents. And something that’s often neglected, we kind of think, oh, you know, here’s an agent. Let’s just, it’s really, really smart. Let’s just unleash it and it’ll work with other agents and things will just somehow magically work out. And that doesn’t happen. And so I think those are important points. The way I usually talk about this is let’s design effectively a nucleus that we know to work, as well as a process for adding new agents that are registered into sort of a trust layer. And I think that would be a mitigating factor towards some of the pitfalls that you mentioned.
David Cushman — Executive Research Leader, HFS Research[05:32]
So is this some of the response to one of our other conclusions in the paper is that while multi-agent adoption is happening fast, enterprises do hit this complexity ceiling without an agent OS, an agent operating system. So that layer that you’re talking about is a standardized layer for governance, orchestration, observability, data trust, workforce integration. So what does this agent operating system mean to you operationally? What are the non-negotiable components you would insist on before attempting to scale with agents across the business?
Babak Hodjat — Chief AI Officer, Cognizant[06:15]
Yeah, I would like for sure to know, to be able to register every agent, have access, full access to the logs of every agent as they’re operating so that we can real time safeguard the agentic system. We will have to work on what trust means. Trust is a big word. And what does it mean organizationally? It’s only there that we can safeguard against emergent behavior. You mentioned emergence. It’s an important concept. And in order to be able to safeguard against untrustworthy or emergent behavior that we might not really appreciate, we have to have a top-down approach there. So I think that those are some of the concepts there that are important. Sandboxing, in some cases, would be very important. Gradual rollout, rather than just, you know, sending these systems out there. Also gradual fine-tuning of these systems. So you have a system that might be using some form of commercial large language model behind the agent, and you’re collecting the data and you do need some human sort of labeled sort of overview of every agent’s behavior and that labeled information, that labeled data of its behavior out in, you know, circumstances that you haven’t planned for before can be used to then fine-tune the behavior of the agents. So one misconception with agentic systems is that they are, it’s a one and done. You just take the best large language model based agent out there and you just unleash it. And that’s just like far from the truth, very, very unsafe and dangerous. So we have to view it as a continuum where we bring these systems in, we orchestrate them, we safeguard them, we register them with a trust layer. And we gradually sort of roll them out. We gradually fine-tune them, make them better, and engineer and re-engineer them. And also define a process, as I said, for adding new agents to the fold. Because, you know, agentification is like peeling an onion. You want to keep adding to it, obviously. And so we need a process for that, for sure.
David Cushman — Executive Research Leader, HFS Research[08:57]
Yeah, it seems to me that there’s, particularly on the governance side, it’s not. We need to think of governance slightly differently than we have done in the past with software, for example. This isn’t software in the same kind of world. What I imagine is the challenge is you want to get the generative capability of a generative AI model. You wanted to have the ability to come up with new ideations and new ways of doing things. That’s its value. That’s where it will come up with something that adds value to what you do. But you want to be able to control the actions, and that’s the agentic side, so they can’t just wreck your database, for example, or book you on that flight that never existed. I’ve seen lots of examples of that. So are you having to shift people’s thinking about governance when you’re having these conversations with the enterprise?
Babak Hodjat — Chief AI Officer, Cognizant[09:53]
Yeah, I mean, first of all, it forces us to think more non-ambiguously about what the guardrails and how we do the governance. It also forces us to think of governance not as this single point of audit that we do, and then it’s a checkwork. It’s something that we’ve done in the past with software or with processes and so forth. And now it has to be an ongoing, as you very rightly pointed out, the whole point of using these agents is allowing them to autonomously make decisions in situations that they’ve never encountered before. And that’s part of their power. But in order to be able to trust that this system will be able to make the right choices, we will need to define what we mean by governance and to audit it on an ongoing basis. The good news is the same agents or agentic technology will give us the means to do that in a non-intrusive way. So you can actually have agents checking the work of other agents without actually interfering with them. But then the question becomes, how do you non-ambiguously define what you mean by trust on an ongoing basis?
David Cushman — Executive Research Leader, HFS Research[11:15]
And for me, that’s how you actually ladder up towards fuller autonomy is to have that ongoing governance process that’s allowing more autonomy as it’s checked and is approved. And you go through that ladder.
Babak Hodjat — Chief AI Officer, Cognizant[11:30]
Yeah, I just wanted to add that like full autonomy might be a little bit of a bridge too far, but it is a continuum as you said like getting more autonomy getting humans to be sort of pulled in where necessary versus rubber stamping every decision that’s kind of where we want to take it.
David Cushman — Executive Research Leader, HFS Research[11:47]
So I’ve got one kind of final question and it’s, I wanted to get into OpenClaw a little bit, because it’s, yeah, it’s the toast of the town around these parts. I’m in the Bay Area myself at the moment. I was at an event yesterday that had been hastily assembled as an OpenClaw specific event in an OpenClaw specific space over in Oakland. Hundreds turned out on a Wednesday lunchtime, which is quite extraordinary, really, the focus on it. The reason why it had some context here, I think, is that in our paper, we noted that there was a lot of talk about MCP becoming the default agent-to-agent coordinator, even though we found that enterprise adoption in the data was only 8% of MCP at that point, with custom-built orchestration at 31%. And so proprietary approaches were still very dominant. What’s your read on why interoperability is still lagging in the enterprise at least? And what has to happen for the multi-agent ecosystems to actually work across vendors and platforms, and which raises for me the question of what might the impact of something like OpenClaw be, given this on-machine approach, which offers a lot more control from the get-go. But obviously, we’ve identified lots of risks along that journey too. So what’s your take on where we are with interoperability and this idea of people being able to use agents from multiple sources to get jobs done?
Babak Hodjat — Chief AI Officer, Cognizant[13:24]
Yeah. The world is demanding certain standards and protocols. And when something comes out, it gets adopted very quickly and sometimes adopted for use in areas that it might not be very well suited. You rightly point out that the report shows that MCP still doesn’t have the same level of adoption as it probably deserves. MCP for, you know, API or app integration for agents is a reasonable protocol. And its use is growing. I think some of what we see in the report is legacy in the fact that, you know, MCP was either not there yet, available through the app, or people just chose to do the integration themselves. It’s not very hard to do that. You know, the world can live without MCP, but MCP gives us this standard that’s very useful. So for that use itself, I think MCP has matured somehow, somewhat to the point where it’s useful, but it’s overused. Like MCP is now being used for agent-to-agent communication. It wasn’t really designed for that. But it’s for the lack of other better designed, more mature sort of protocols for agent to agent that people are kind of reverting to MCP and MCP is kind of lacking there, too. We see that a lot. Then along comes OpenClaw and, you know, the whole skills kind of paradigm, which is powerful, yet, you know, for example, security and safety were not top of mind when it was designed. And so it opens you up to all sorts of risks. So, yeah, we’re kind of eager. I mean, I can get that. As an engineer, I get it. Like you’re eager to take the next new powerful thing that enables you and empowers you to do something that you weren’t able to do that easily before. And that kind of explains the fast adoption. But what’s lagging, which is scary, I think somewhat, is the security thinking of this. And the fact that this is not just being deployed within controlled environments or for games or toys or whatever, toy problems. They’re actually out there. And in some cases, they could even carry your wallet around and do transactions online, for example, or fake identity. The whole agent identity concept has not been resolved. There’s a lot of threads, a lot of tracks that are looking into this. And it’s yet another one of these areas where there’s a lot of demand for something to become standard, but we don’t have anything just yet. And I think, you know, piggybacking on protocols or standards that were really not designed for this or not very well thought out or kind of patched to include these types of features, you know, are, I wouldn’t say a disaster waiting to happen, but, you know, they open you up to risks, right? So, yeah, I think we’re living in that kind of a world right now a little bit. And, you know, David, the fact is these things are unfolding at such rapid pace. You know, we just haven’t had time to, you know, take a step back and think about safety and think about, you know, enterprise grade multi-agency as much as we should. And my hope is that this doesn’t come on the back of some sort of major issue that weakens the whole momentum here. I’m hoping that we can keep up with this very high demand. I mean, people are working on this, but, you know, it takes a while. It’s kind of safety and trust. Those are kind of lagging a little bit.
David Cushman — Executive Research Leader, HFS Research[17:52]
Thank you. I think my takeaway is that the lessons you’re looking for, the kind of guidance to the market, a lot of that is encapsulated in this paper that we’ve just published. There is some good guidance on how people should approach their scaling of agentic and multi-agent systems across their own enterprises. Some pauses for thought, some evidence from people who’ve gone down that journey and crashed into some of the barriers and the ways in which we can resolve that. So I’m hoping that it’s going to do a lot of good in the market and actually accelerate the uptake of agentic because we all need this decoupling, right? We all need this capability to grow without necessarily adding headcount and complexity to organizations. That’s the margin killer. And this is a real opportunity for every enterprise, really, to take a step forward to what we’d describe as the Services-as-Software future. Babak, any final thoughts from you?
Babak Hodjat — Chief AI Officer, Cognizant[18:53]
Yeah, I think the paper is a really good indicator. I was somewhat surprised as we were doing the research together as to the scale of adoption, which is a harbinger of widespread use beyond the standard areas. So I would encourage everyone to take a look at the paper. A lot of very interesting sort of information there as to the state. And I think even some advice and, you know, some knowledge as to how to do this the right way emerges from the paper itself. So hopefully people will find it useful.
David Cushman — Executive Research Leader, HFS Research[19:36]
Babak, thank you so much for your time today. Have a great day. Thank you.
