Swapnil Bhartiya: When it comes to AI, everyone assumes that AI race is all about models. But teams who are shipping agents into production know the reality. There are a lot of other factors that play a very big role when something goes wrong. What breaks is not essentially the model, but things around it. Orchestration, session state, sandboxing, memory, tool, routing, essentially the harness. And most teams are hand rolling that harness themselves, rewriting it every few moments as new models improve underneath it. Anthropic know that, and that’s why they have quietly shipped something called cloud Managed Agents. It’s a fully managed agent runtime that decouples the brain from the hands. And the implications of how enterprises are going to build agents and govern AI are much bigger than the noise that this launch created. Today we have with us Amit Nayak, VP of AI at CData, who has been digging into exactly what that means and what companies gain or lose when they hand that harness to anthropic. Amit, it’s good to have you on the show.

Amit Naik: Thank you so much Swapnil. I love the introduction that you just gave it and I wrote a very detailed blog post on this and I’ve been looking at this problem for a long time.

Swapnil Bhartiya: Before we get into the meta harness idea, let’s set the stage for our audience. First of all, of course we have hosted you folks so many times, so our audience know, but I want to talk about CDATA in this context. So talk a bit about what CDATA does and what’s your role there and what has been keeping you awake at night and getting you excited every morning.

Amit Naik: Thank you so much for having me on the show. My name is Amit Naik. I’m the VP of AI at CData. CData builds the infrastructure for AI agents and humans to talk to different enterprise data sources. We’ve been doing that for the past decade or so. We have a battle tested range of more than 350 different connections. My role at CDATA is to focus on the platforms and the systems that we use to build AI. Primarily our Connect AI product, which is a managed MCP server and pretty soon having interfaces for CLIs and different places like that. The part about what keeps me awake at night and what gets me excited is all rolled into one. Things are moving so fast and moving so rapidly that we are in an unprecedented age of experimentation, while at the same time we have the challenges of making sure that a lot of those experiments translate well in production. What might work in a demo, which gets engineers excited, might not work well or might even fail miserably in production because of the unique nature of some of the constraints that come. When you are talking about production systems.

Swapnil Bhartiya: We use the term meta harness. It’s not a buzzword. It is a very specific, you know, you can say term, but what does it actually mean? And why do you think that it matters more than improvements in the model themselves? Here’s the problem with models. New models keep coming out. That is also not a big issue. But the layer these companies put on top of front end models. So your prompt, I mean what I have been struggling lately is that cloud will not even follow the prompt. It will, no matter how neatly you have written it. So organizations cannot keep redoing it. So we do need something. The word harness is a very strong word. So I want to understand from you, what do you mean by meta harness? What is the concept behind it and what does it actually do to kind of rein these models?

Amit Naik: Sure. I think extremely interesting angle to attack the AI understanding and the AI systems concepts. So let’s take a step back. Typically when people interact with AI systems, they think about ChatGPT, they think about Claude. And what people assume is that they are interacting with a large language model. The large language model is only one part of the system. They are actually talking to the entire system systems architecture around it. Where if you have an agent, the agent is composed of a few things. One is the user interface. The second is the LLM, which is the brain. And then you have different skills and tools which allow the agent to interact with the external environment. So typically the way an agent works is you receive a user prompt or a trigger. The LLM reasons over uses skills and tools to interact with the external world. It receives feedback on how that interaction went and it goes in a loop over and over and over until the prompt that it received made sense. The task is achieved, the goal is achieved. As you can see, LLM is only a small part. There are systems engineering software around all of these which include things like managing memory. You don’t want the LLM to forget between two requests what the original prompt was. There is session maintenance, there is how to use tools. There are things like guardrails, there is observability. All of these components together are what make a harness, which is what makes an agent an agent. Essentially, if we take one step above this, you have a meta harness. Why does that meta harness even matter? You know you have the harness. So okay, what’s the big deal? Why do you even need a meta hardness? It again goes back to the fact that the model capabilities are evolving so fast that every two or three months companies have to rebuild the way that they manage memory. Is it short term, is it episodic, is it long term? The way that they manage the tooling, the guardrails, the various systems around it, these are seeing tremendous changes. As a user, you can’t keep changing the way you interact with these systems every two or three months. This is where the concept of a meta harness comes into play. The metaharness is what the user interacts with. Inside the metaharness you have the actual harness which undergoes rapid change because of model improvement capabilities. Because what these frontier labs do is they add more and more capabilities to the models that is absorbed by the harness. And the meta harness gives the user a level of stability with which they can interact.

Swapnil Bhartiya: Where are teams getting stuck? Where they are hitting the wall when they have built their whole agent and they now want to move from demo to production ready agent because that’s everything works in pilot and most actually pilots also fail. But reaching production is where the challenge happen. What tends to break, where we do need to bring in meta harness.

Amit Naik: I think this is a very familiar problem with a lot of demo to production translation that happens which software engineers very frequently deal with, but that has become supercharged with AI because there is so much hype. Actual innovation, model capability, improvement, all mixed in. POCs are actually very simple to create. Now the bar has not never been lower. You can just ask Claude Code or Codex to actually spin up a POC for you. Where things get sticky, like you rightly pointed out, is when we try to translate those POCs into actual production systems. And there are three main areas where I’ve seen friction come in. I’ve been in this space for more than 10 years. The last three of them have been exclusively around generative AI systems. I’ve worked at different companies where I’ve brought multiple first iterations to life. So I think I’m pretty uniquely qualified to speak on this. And there are three main aspects where things go wrong. The first one is context management. In a POC, because time is short, because you want to get the most bang for the buck, you give it simple prompts, you give it a few turns. So the context of the LLM is not polluted in production systems. Users may go for many, many minutes, sometimes hours that the system is running. If you just think about all the information that starts to build up in the context of the LLM, there is high chance that things will go wrong. So you might have seen that for the maybe last six months, maybe a year. Context engineering has become a major part of the discourse because people have moved from prompt engineering to making sure that the right information stays in the context of the LLM. We have problems like context bloat, where lot of unnecessary junk has been dumped in. We have context rot. Maybe some of the information has become outdated. So the agent is going wrong, not because the LLM is failing, but because it is acting on outdated information. So this is one aspect. The second important aspect is the LLM is mostly the reasoning or the thinking layer. But what gives it agentic abilities are the tools that it uses to interact with external systems. In a demo we don’t use too many tools or the tools that we use have been guided along certain predetermined paths because it’s a demo. But in real production systems, you need the ability for the tools to reach into enterprise knowledge sources. That means you need good authentication, you need authorization, you need traceability, you need the right output to come back from the tool. And in real enterprise scenarios, these things are very hard to get right. So what worked in a demo pretty seamlessly because there was one record in a database. As soon as you go to an enterprise system with millions of records, hundreds of schemas, that’s not going to work anymore. The third one is how often do you see in a demo observability or auditability? Almost nobody talks about these aspects. That’s not what a sexy demo means. You are usually showing off functionality or showing off features. But in an enterprise system, just imagine your security group, your IT system administrator, they are trying to make sure that you have auditability, you have traceability. You need to know which agent accessed what. What happens if the agent suddenly goes rogue and starts deleting a bunch of data? You want governance. All of these make translating the slick demos into actual production pretty hard.

Swapnil Bhartiya: When we look at managed services, or if you look at, you know, Anthropic’s managed platform, of course any managed service is open edit, right? It had made a decisions for you. So it does from where we look, it does take a lot of that operational burden off team’s hand. As you talked about the problem challenges, but it also comes with a lot of trade offs. You do lose a lot of control flexibility. Can you talk about from your perspective, what do companies actually gain from this approach and what do they lose? And if I ask you frankly, whether the trade off is worth losing the control?

Amit Naik: I think great question. Speaks to the heart of the decisions that executives and architects have to make. Evaluating the Managed agent space. If you take a small step back, I think agents come in two flavors. One is desktop agents. These are things like Claude code codecs which work on your personal desktops. They assume your identity and they go away the minute you kind of close your laptop. When we talk about managed agents, these agents are running in somebody else’s, the provider’s infrastructure, which is what makes them cloud agents. And second, the provider is actually taking the burden of all of the harness engineering that we talked about. They are managing the memory, they are managing the sandbox, the tools, the tool outputs. So all of that operational burden is on the provider. Which is where I think if we look at what are the advantages? The advantages are that you are freed from that managing operational burden. Your company may not be the topmost expert at building these distributed systems. Your core business, your core business, your core IP might be completely something else. You don’t want to take on the operational burden of building this harness, managing memory, making sure the tool calls work. This is a distributed system. So the primary advantage is speed. Cloud managed agents allows you to build agents very quickly. The configuration is simple, a few clicks. If you want a no code environment, low code environment, you can configure a few data sources. In a few clicks you have an agent which is ready to. So speed becomes a very critical factor. And in today’s fast moving AI age, speed is a great muscle to have. If you can go from your idea to the finished product very quickly using cloud managed agents, that gives you a tremendous business advantage. So that is advantage number one. Second, the ongoing maintenance of it is also now outsourced to somebody else. It is outsourced to cloud, it is outsourced to Anthropic. You don’t have to keep managing like the memory, the harness, the tool call, the different aspects of those. If model capabilities suddenly take a leap, let’s say you have Claude Mythos, tomorrow you might have some other model and the harness needs to change. Well, that’s Anthropic’s problem because now you’re using a managed product. The third one, which is also kind of related here, is that the default might be set. Well. Anthropic, one assumes, knows the product well. So they have set the defaults in the harness. How many days to keep memory, how to do tool calls, how many turns to take, which sessions to surface. All of these are default. You assume that the defaults work well. So these are some of the benefits. Balancing off against that, what do you lose? Primarily, you become vendor locked in at some stage you have kind of bet that Anthropic is what I’m going to configure the agents in. And tomorrow, if OpenAI or Gemini or one of the other providers comes up with a better model, well, you now have to either replicate what you did here in their system or you are kind of left out of that model improvement. That’s one problem. The second problem is you are introducing runtime dependencies. Now, let’s say you have some very critical business process and you have automated some portions of that using Claude managed agents. Your uptime is now completely dependent on Anthropic’s uptime because it’s using that agent. And if this is a very core part of your business function, of your revenue, every outage at Anthropic’s managed agent service is an outage that you face a revenue loss that you are taking. So you have to be very careful on which scenarios you use these for. The third one is something which is underappreciated by businesses for now, but that is how an agent works when it approaches certain systems, how it puts together the information is your business. It is essentially a software description of how your business functions. That is a very core IP that is now in somebody else’s system. Are you comfortable with that? Is that something that you are okay with? If yes. And to their credit, some of these providers, even Anthropic has said that you can take all of those and take it with you if you want. You can export it. So maybe that is not a big concern, but that is something that people should be thinking about. Finally, there’s always the cost angle, building it yourself versus cutting somebody a check. Where does the ROI kind of make sense? If you have a large engineering team which has expertise or which knows how to build it, maybe it’s cheaper to build it in house rather than cutting Anthropic a check. On the other hand, if you are not an expert at that, maybe it is actually better to cut Anthropic the check. So revenue considerations will also come into play.

Swapnil Bhartiya: I want to talk about because CDATA data layer I will talk about as all of this mature, what does it mean for data layer? What does it mean for governance, real time context, access control? Where does all of that fit in the agent stack? And because there are a lot of industries which are regulated, they have to do a lot of things with compliance. Where do you see C data in this scenario?

Amit Naik: Yeah, I think extremely, I think complementary question here, which is that if the meta harness is controlling all of the agentic reasoning. The complement to that is a data layer. Those two have to work together. One can’t exist without the other, without having a pure data layer. But no, reasoning is not going to work and only reasoning without data is not going to work. So they are kind of complementary with each other. And this is where I think I see cdata playing a key role which which is that you have the meta harness which acts as the reasoning layer or some of the abstractions around thinking tool use. But that needs to be complemented with a data layer which exposes all of your enterprise data where all the action is to the agents in the harness. So CDATA plays a key role there by exposing all of your enterprise knowledge sources in a governed way. So providing connectivity to more than 350 different sources, like I mentioned previously, but also creating context. You don’t want your agents struggling to map between two different data sources. How does a field in Salesforce compare to another field in Jira, for example? And CDATA makes those semantic connections easy. Finally you have the topic of governance. This is where I think CDATA plays a key role. You have different ways of authenticating against different data sources that gets pushed down and inherits the access controls that you might have on any of those sources so that you don’t have to redefine those. And then you have different interfaces which make full auditability possible. So your agents can have governed semantic and connectivity access to the underlying sources. Which is where CDATA really shines as a data layer for all your agent needs.

Swapnil Bhartiya: What else do organizations lose through manage agent? Let’s talk about because you mentioned memory traces a lot of other. What I like about mcp, not that I am against people, but one of the risk is that most time people all of their knowledge in their head when they walk out of your office that knowledge goes with them. Individual knowledge through you know, of course memories, all those things. With AI that knowledge becomes the integral knowledge of the organization itself. Anybody can tap into that. When we move to these managed agent, what happens to that? You know, institutional knowledge because now there is an external provider that owns the runtime, not you. How should enterprise think about protecting their institutional knowledge, institutional intelligence as they adopt these managed platforms or you feel that’s not a big worry at all?

Amit Naik: No, I think this is where the discussion will move. Once people start uptaking agents, especially in a managed scenario, what happens typically is first enterprises try to automate certain processes. This is where the institutional knowledge starts to get defined. The process of creating that automation is what brings something that is in one person or a group’s individual memory into what is institutional memory. If you take this a step further, agents are now the ones who are executing this automation. And if you use managed agents, it means that all of this execution of automation is happening in somebody else’s infrastructure. This is going to become a significant challenge of how do you capture it? Agents have sessions, they have traces, there are open telemetry standards around how to capture some of those traces. And enterprises need to invest in making sure that they have full visibility into these traces and they have the ability to kind of export all of these traces to capture all those automation IP that they have, which is their core business. The other part to remember is not to move critical data assets onto the managed platforms. For example, if you have core databases, if you have core knowledge sources, it is best to use an MCP server. But keep that out of the managed agents, only let the managed agents have governed access to it. That is the other part of the equation. So you have this institutional automation which you capture using agents and then you have the institutional data to which you are giving the agents governed access. So with those two, you can mitigate some parts of it, but as soon as you adopt like a managed agents platform, you’ve kind of opened the door towards at least some IP being transferred to the hosting company.

Swapnil Bhartiya: So I think at some point we’ll have to just accept it as a kind of trade off. But what value does it bring where you see that? Yes, even if you have to make those trade off, it is freeing so much of your time, so much of your resources where your teams are not. I mean, in the traditional IT sense, we spend most of our resources in plumbing, in keeping the machines running, keeping the pipes running versus taking actual shower, if I can use the analogy. So here also the same thing is happening. Your teams are saving so much time on that now you have to think about it strategically. So if you are advising a team today to shift fast on a managed harness on or build on their own, what is your honest answer? Where you weigh all the pros and cons and you’re like, this is what you should do.

Amit Naik: I think this is a question that will depend heavily on two things. One, what is the maturity that you have for running AI? First, processes like ages. If you are at medium to high levels of maturity, then it actually makes sense to try and manage all of this yourself. But fairly few enterprises are at medium to high levels of maturity. Most enterprises are at the beginning phases of the agentic revolution. Given that I would say, rather than burn your fingers with trying to set up large infrastructure investments or large software investments, use a managed agents platform and a managed data platform. The two go hand in hand together. Focus on the automation aspects more, the people aspects. Figuring out how people execute certain processes, how they can be automated, and how they can be moved to these agentic platforms. That should be your lowest friction. Fast adoption curve where you will begin to see rapid benefits, ROI, positive ROI. You know, if you want to now start owning certain pieces of the managed infrastructure that might be your logical next step, you might say, yes, I will use the entirety of the managed platform, except for maybe regularly storing the memory back into my enterprise so that all of the institutional agent traces that you have are now captured even in your enterprise. So tomorrow that frees you up to experiment with. Okay, OpenAI has a better model, Google has a better model. Let’s see how their harness works. You have all your institutional memories captured. The final step might be okay, now we understand how our AI maturity functions from a people perspective, from a team perspective, from a hardware software perspective. That’s when you should even possibly try to attempt to build an agentic infrastructure. Until then, use managed stuff.

Swapnil Bhartiya: Amit, thank you so much. It’s really, I mean the way you look at it, of course Metaharness is very sharp vision at where the industry is moving, where everybody else is heading. Thanks for bringing that clarity, that balance and of course folks who are watching it, please go and check out CData, especially if you are building agents and wondering how your data layer needs to evolve. Once again, thank you Amit and I look forward to chat with you again.

Amit Naik: Thank you so much for having me Swapnil. Thank you.