Name: Mohak Sharma, HoneyHive
Uploaded: 2025-04-25T17:22:00.000Z
Duration: 25 min 6 s

Clips
More from theCUBE + NYSE Wired: AI Agent Conference

Mohak Sharma

Co-Founder & CEO

HoneyHive

play_circle_outline HoneyHive's focus on AI-native enterprise solutions and DevSecOps in AI

play_circle_outline Founders' background and origin story of HoneyHive

play_circle_outline Comparison of cloud-native development to AI infrastructure

play_circle_outline Shift towards autonomous agents like developers in CI/CD pipeline

play_circle_outline Advice for enterprises on getting started with AI agents and experimentation.

Info
Transcript

Mohak Sharma, HoneyHive

Mohak Sharma

Co-Founder & CEO HoneyHive

In this insightful episode of theCUBE, we welcome Mohak Sharma, co-founder and Chief Executive Officer of HoneyHive. Sharma joins us from New York to discuss the AI Agent Conference 2025, held in collaboration with NYSE Wired. The conversation explores HoneyHive's pioneering work in observability and instrumentation for artificial intelligence agents, examining the significant opportunities AI presents in transforming enterprise frameworks into AI-native ecosystems.

John Furrier, co-founder and Co-Chief Executive Officer of SiliconANGLE Media, engages ... Read more

explore Keep Exploring

What was the inspiration behind starting your business in the AI industry? add

What is the origin story behind the creation of the company and its focus on observability and background in Generative AI systems? add

What are some of the similarities between the evolution of agentic systems and cloud computing in terms of infrastructure build-out and distributed systems challenges? add

What role do AI agents play in driving improvements in companies' infrastructure and operations, particularly within the context of DevOps and security? add

What are some key tips for implementing AI technology within a company? add

bolt Powered by CUBE AI

Mohak Sharma, HoneyHive

>> Hello everyone. I'm John Furrier, host of theCUBE. Here in our Palo Alto Studios, we have remote coverage coming in from New York City as part of the AI Agent Conference support, but also unpacking what's going on, this agent infrastructure layers around the stack where the innovation is that's going to enable and empower this agent technology wave. Obviously, it's hyped up beyond all recognition, but there's agents out there, it is happening. It's the real deal. This is not a bubble, but it's just progression. Mohak Sharma is here, the co-founder and CEO of HoneyHive. Mohak, thanks for coming on from New York. Appreciate you coming to our Palo Alto special media event around the AI Agent Conference.

Mohak Sharma

>> Yep. Thanks for having me here, John.

>> So you guys are doing some really cool work around some of the software that's going to enable agents and specifically around kind of observability instrumentation, understanding that is that right? And give me a taste of what you guys are doing in the business. How big are you guys? Where are you at? The originator story. Feel free to weigh in, spend a couple of minutes, talk about what you do.

Mohak Sharma

>> Yep, totally. We started in the business roughly about two years ago actually with the general idea that every enterprise over the next five to 10 years is going to be an AI-native enterprise. I think the big opportunity that we saw was what does the DevSecOps lifecycle in this AI agentic world really look like, right? I think the big issues that we really came across is AI's quality problem, right? If you're trying to build some sort of an AI agent today, it's very easy to prototype it, but actually building something which is reliable, robust, and really safe, right? That's hard. So I think one of the big things that we wanted to crack was exactly how can we test these systems effectively, how can we observe these systems, how can we actually explain their reasoning thoughts, their traces, so on and so forth, so developers can actually know what's happening, understand where they're failing, and continuously improve these systems. The origin story is interesting. My co-founder and I, we met at Columbia, we were roommates there. Dhruv was working at Microsoft after graduating building Office 365 observability, so that's the observability focus and background where that came from. I was working on data platforms and systems at a series D enterprise startup here in New York called Templafy. And at the time, both of us started building agentic systems ourselves. This was summer of 2022, so still very early. I think Generative AI was not even coined as a term back then, and that's when we realized I think every enterprise is going to be building these systems, yet quality measurement and safety is a huge problem in this entire space. That's what led to the story.

>> Well, first of all, I love the observability angle because we love observability. We've been covering that on the cloud-native side, and if you look at the market, I want to get your thoughts on this because there's been a lot of discussion around what's the TAM? Is it real? I mean, first of all, I think agents are totally real. They're definitely out there now happening. They're not just chatbots, but there is a lot of headroom in the stack for full agentic autonomous opportunities. But if you look at the wave of cloud, cloud-native, spawn, the SaaS generation, now the agents are coming and we are predicting that the agent market will look a lot like SaaS but bigger because the enterprise side can actually not just build an app but actually be like a SaaS company like Dropbox was or Airbnb. They can actually have a huge advantage. So the thought is, okay, cloud-native, we saw the rise of the CNCF at Linux Foundation, observability played well at scale around cloud-native services, microservices. That piece of it was huge, and look at where we are now with say, Kubernetes. So as you project that out to say AI infrastructure, what's your vision on this? Because I see a correlation between what data is doing and what cloud services are doing. It's kind of in the weeds, but I want to get your thoughts and reaction. One, do you see the same picture? And if so, or if not, what is happening? Because you're seeing data engineering to address the scale problem. Because you're right, it's easy to do a prototype, Hey, look at it, but now run it across the enterprise. Good luck with that. I mean, it's not easy. What do you think?

Mohak Sharma

>> Yeah, totally. I think one of the biggest lessons from this wave is agentic systems, I consider them the holy grail of productivity and automation. I think cloud really what we are trying to do with productivity software to a large degree is automate 10 to 20% efficiency within enterprise workloads. I think agentic systems have the capability to push that bar to roughly that 80 to 100% automation, which results in a lot more enterprise value, which is being created. That's it. I think some of the similarities where they lie is very similar to cloud. We are in this journey where we are trying to build infrastructure from the ground up. There's a huge infrastructure build out which is happening first of all, I think in the cloud days we are building data centers housed with a lot of CPUs, EC2 with the name of the game. I think today with the GPU build-out, a very similar data is happening, right? Every single company is trying to become an AI-native enterprise. Now, what that results in is similar distributed systems challenges. If you're building an agentic system, which can let's say take actions within Salesforce and be able to automate some sort of a workflow, the same problems of how do you trace the system and the traceability of what happens when the user gives this query? What happens when this data flows through container one to container two, and then there's the semantic reasoning, which is also happening at the same time within the tokens of these models. How do you trace that? How do you monitor that? How do you evaluate that? The DevOps challenges don't change. I think it was a very similar sort of trend we saw with cloud where all systems were on-prem, we move to the cloud, the distributed explosion happened, which necessitated the need for Datadog and the likes to come in and essentially provide telemetry over that data. I think we are going through a very similar wave today.

>> I mean that DevOps angle is amazing because I think I agree with you because if you look at DevOps, Dev was the cloud, right? DevOps together, I mean, born-in-the-cloud companies didn't have IT departments, okay? If you go to the enterprise where they do have IT departments, then the Jensen Wong phrase of IT is the HR department for agents, that over-the-top haymaker he threw out there is interesting because in a way, he's saying the same thing as what DevOps did for IT then. I mean, any company that was born in the cloud really didn't have an IT department until they got too big, and then that's happening in agents. So take me through that because end-to-end workflows and domain-specific data is where the success in AI is, but that's also the challenge on the DevOps side. So is there an AI ops redefinition here? Is it agent ops? Talk about that operational engineering piece and of course, for fun, if you want to throw second there too, you can throw that in there because DevSecOps, which translates to AI DevSecOps too.

Mohak Sharma

>> Yes, yes, totally. I think initially when this industry started in 2022, the entire category was called LLM ops, and I think we've been through a couple of iterations really right at agent ops, which I think captures the entire ecosystem. I think when we think about the operational complexity, one of the things that pre-training has offered us is that these models are now available to be used by an API, which makes it very easy, very accessible, but then prompt engineering is a completely net new category which is being built on top of these models. How do you provide the right context, the right instructions to these models in order to be able to adhere them and tune them in a specific way? I think there's a huge set of tooling which is now built just to be able to support that new workflow where coding and instructions are essentially in natural language in English. So that's one of the key areas of the stack which is being born. The other area, which is very interesting to us is the orchestration layer. All the agentic frameworks like LangChain, LlamaIndex, the of the world. This layer is essentially trying to piece together how can we work with multiple models that exist within the stack and piece together a workflow. They can actually automate something meaningful. So that's the base layer upon which the entire foundation is getting built now. On top of that, we see observability to be able to see exactly what's happening in the stock, and there's the net new category which is emerging now called evaluations. Evaluations, what we realized is really the core piece behind most of these real-world agentic systems. How do you measure these systems? Accuracy, quality, performance, how do you improve it? How do you know if you're actually improving or regressing? Every time a new model comes out by OpenAI, how do you compare if that model is actually better than the other ones that exist out there? So I think that piece of the stack is going to be extremely important when we think about this entire ecosystem, and that's where our company is really focused on. Some of the other areas that we've also seen are fine-tuning GPU providers further down the stack where there's a lot of interesting innovation happening, but closer to the application layer, I think that's where we've seen the most interesting stuff really happen.

>> Yeah, it's interesting. We've been saying on theCUBE for years, platform engineering and data ... we call it data engineering back in 2020 because we saw that SRE vibe going into data, all the top people that were doing the best data management work were thinking horizontal, but now you got the domain-specific AI where the agents shine. It really is a platform engineering, DevOps culture. Moving into agentic coding where you're at, do you see that same thing? I mean, because it's an extension, it's not like a replacement. I mean you're looking at all those same frameworks and paradigms of cloud-native moving into AI because that sets the table because that enables. So do you agree and share your thoughts on that, and then two, what's that enable because it's a disruptive enabler on top of the growth of say, cloud-native?

Mohak Sharma

>> Yes, I do agree to large degree AgentOps really is like a derivative of DataOps and DevOps to large degree. I think the biggest change that we have noticed though is the personas which are actually involved in this sort of operational process. Typically, if you think about DevOps, really it's engineers, SREs. People with technical skills were involved in maintaining these systems, debugging them. I think today when you think about the AgentOps world, you don't have just developers and MLEs, but you also have domain experts, right? Subject matter experts who understand, let's say how a financial agent should behave or think about a copywriter who might have thoughts about how an agent's blog post should be written. I think those people are increasingly relevant in today's world, which was not the case in DevOps and DataOps. That's where we've seen this huge difference between the two approaches. The other difference I have-

>> Go ahead.

Mohak Sharma

>> Just like a closing argument, the other difference I have to really point out is the iterative nature of agent engineering. I think when you think about data and cloud, it's deterministic software. The code compiles exactly how it's written. You plan something, you build it, you test it, you put it in production, and you can forget about it to a large degree. I think agents, while you can build something in a very short amount of time and test it maybe against 50 test cases, you put in production and then you see the real long tail of user inputs and that's where it actually breaks, right? We've seen incidents like Air Canada's chatbot, Chevrolet chatbot, which was very popular roughly about two years ago. Those sort of issues happen in the real world, so you need to go back into the development phase, continuously improve your model, and I think that's where the key difference lies between the two approaches.

>> Yeah, great insight, Mohak. I want just go one step further because one of the things that we've changed our opinion on in theCUBE and theCUBE Research is if you go back to 2017 through say 2021 or 2020, the observation on our side was, oh yeah, platform engineering is great as we talked about, but the data people, the data departments are going to look like security departments. They're going to set rules and then shift left in the pipeline. What you're getting at is much more in the CI/CD pipeline thinking. Less here's guardrails, although guardrails are talked about in AI, less of a here's the parameters, go, because there's more stuff going on. So our new opinion here is that agents will not only be self-sufficient and autonomous, but they'll be like developers. All those CI/CD pipeline conversations around coding will be what agents will do. That's kind of what we see. Do you agree with that? What would be your perspective? Because that shifts it to looking more like DevOps, less like the security department.

Mohak Sharma

>> Yes, exactly. I think that's very right, and I think I would just go a step further and say most of these companies' improvements I was talking about earlier are going to be ultimately driven by AI agents themselves. I think this is the beauty of autonomous intelligence systems that you find issues, you're not relegating it to a developer to go fix them, you have AI agents that can plug into the ecosystem and essentially root cause the issue, fix them, deploy them, and essentially plug into the CI/CD infrastructure itself. So the sort of world we see is this would end up being certainly called the meta agent, where essentially HoneyHive's infrastructure tests these agents. It gives you some level of visibility into what's happening, but AI agents themselves will now take those insights, go ahead and continuously fix and improve your pipeline. I think we are very close to that vision today.

>> Talk about the role of observability in this kind of new world because if you just simplify it and zoom out, you say, okay, I got data over here, I got origination of data, I got some data, what's going on in the workflows and the end-to-end stuff and the services are deployed and then the insights that come out of that, there's kind of this messy middle. It used to be pipelining data and all kinds of transformations. So we think agents will fill that middle layer to your point about them being autonomous, which means that if you don't have the right data, you're not get the right insights, you don't have the right agent architecture and understanding what's happening, observing that, it doesn't really work on the business side because at the end of the day, the business value will be the ultimate transformational impact. Talk about that messy middle between having the data and environmental situation to the insights that are needed to come out and then what observability and what goes on in that middle.

Mohak Sharma

>> Yes, that's a really good question. I think that's the key difference between what observability looks like in the agentic world and the traditional cloud DevOps world. I think you're not just looking at code execution, but the reasoning tokens. The messy middle essentially is what I call national language. It's English. That's where things go wrong. Today let's say you're building an AI agent. You have something like Datadog. The agent executes, there are no errors, you'd think the agent is perfectly fine. Then you look at the agent's outfit and you realize, man, this didn't really answer my question. This didn't really complete the task. Why does that happen? Because it's the agent's internal reasoning in natural language, which is faulty. Something went wrong somewhere. So the observability in this world is not just trying to look at API errors. It's not just trying to look at code execution, but the reasoning tokens which are being emitted by these models and really run algorithms and models over those reasoning tokens to give you a sense of is the reasoning correct, is the model factually accurate, is it hallucinating? Those sort of insights. What bridges that gap is really using large language models to evaluate other large language models. This is a very meta-concept, which has emerged think in the last few years, but really the idea of being agents can evaluate other agents, and so I think that's what really is bridging that gap between the two ecosystems.

>> It's so much fun because what's happening is the technology from the bottom-up is connecting with the business logic and everything in between are things that we kind of know about, like computer science principles, coding, engineering, but it's got a data piece, but it's also got some business policy thoughts as well. It's not a general purpose, so it's a lot of engineering going on. I mean, it seems like an incredible time. So what's your advice to folks out there trying to get into this world and custom deploy agents? I mean, enterprise right now is scratching their heads going, is it a tailwind for me? Is it a headwind for me? How do I evaluate workloads? How do I benchmark? How do I deploy? What's your advice? Because this is really a discovery market right now where people just want to figure out, how do I get started?

Mohak Sharma

>> Yes. I think we are talking to a lot of Fortune 100 enterprises as well who are in a very similar belt. What's worked for them really is a thinking about what are the sort of use cases that AI agents today can actually accomplish, usually, it's something simple that a human takes anywhere from two to five minutes to be able to perform. Summarization, being able to look at a long-form document, answer some questions, maybe fill out a form. Those are very simple use cases that are very reliable today. So my first advice would be create an experimentation culture within your company and try to go out there and think about what are the simple use pieces that AI today can indeed solve, right? Two, think about the business logic. How would a human perform the same task? Essentially what we are trying to do in the agentic world is replicate the same steps human would take in order to complete task A or task B. So being able to really draw out what exactly a human would do in that scenario, what other requirements for that human success in that scenario helps because that's a business logic. You're trying to port over and engineer into your agentic system. And then most importantly, I think start small. A lot of enterprises out there are trying to make it big, automate extremely large, complex workloads, right? AI is fast-evolving, the models are improving at a rapid pace. I think it's better to start small focus on a concentrated use case, really make it reliable, and then write the wave. These models are going to get better in just three months. So I think there's a lot more use cases, a lot more workflows that are going to be unlocked, but I think today what is possible, we need to be very well aware about that and really consider reliability at its core. Because again, if you deploy an agent, it doesn't work. There's more downside risk than the time it's going to save your enterprise.

>> You got to do the platform work. Oh, great stuff. Brian Baumann at NYSC Wired and I were talking with Simon too about the New York tech scene out to theCUBE has a studio there at the NYSC and our studio there on the floor. We'll be doing a ton of coverage. New York City I love right now because one, grew up in New Jersey, so I kind of love New York anyway, but it looks a lot like Silicon Valley. I'll get your reaction to this because on the New York Tech scene, Silicon Valley, if you're a startup, you come here. Why? Because all your customers are here. All other SaaS companies, the VCs are here. You got VCs in New York, and so if you were starting a check company, your first customers were here and you do biz dev deals, all kinds of karetsu's going on, all kinds of networks. In New York, you're there, folks we're interviewing are there. Your customers are like, you can walk three blocks and hit 10 customers. New York is customer-rich for AI because again, enterprises aren't just being a customer of a SaaS company, they are the customer, they are the application. So agents are enabling the enterprises to be their own thing for AI, not being a customer, so to speak. They can build their own. say their apps, but they build all the data. So talk about the New York scene. Do I get that right? A lot of customers there for you, what's it like there? Share the vibe on what's happening in the New York tech scene.

Mohak Sharma

>> Yes, totally. I think this was one of the biggest reasons we wanted to base our company in New York City as well. I think majority of the companies which are driving the most value out of the tech systems, rightly so, are enterprises, right? More enterprises today are building these tech systems than they're buying tools. So that's a really interesting shift that we've noticed compared to let's say the SaaS era where really most companies prefer to build business about. What that also means for us is we can walk a couple of blocks in midtown, go to the New York Stock Exchange, or go to Morgan Stanley or any of the banks that exist, hedge funds that exist. These are some of the most, let's say, cutting-edge institutions today, which are implementing these agenda systems, deriving a lot of value, and I think having this customer base within the city really helps because you can meet people in person, really be able to communicate what you're doing, understand their pain points and work together very closely. I think the other really interesting trend is that from a VC ecosystem perspective, I think New York is second to only San Francisco, but the number of AI startups and the number of deals that we are seeing within the ecosystem has been steadily increasing. We have some really great infrastructure companies based in the city as well. Datadog is a really good example within the observability space. Starter in New York grew in New York. Within the agentic system, I think we have companies like Crew AI, , Norm AI, who are doing some really interesting stuff. So having all of these customers over here, all these companies being born here also means that a lot of VC dollars are being allocated towards this ecosystem as well.

>> Yeah, and the interesting thing too is that with the digital technologies becoming first party with face-to-face, I mean I'm sure you have tons of friends in the Bay Area, as does everybody.

Mohak Sharma

>> We do.

>> So you're seeing this connection, this highly frictionless communication vehicle. You can just do a Zoom, fly out, get on a plane. So I think it's kind of becoming a one thing, right? It's like Silicon Valley and New York. I mean, it's not about which one's better. They're all kind of connected now. So it's like one thing, and I think that is the interesting thing we're seeing in New York is that the enterprises are driving a lot of the value. They are progressive customers and they have an imperative to be successful and they can justify it. It's not like a POC farm. They actually have urgent needs.

Mohak Sharma

>> Exactly. Exactly. No, I think that's a really good point that you made as well, the digital economy. I think just having Zoom and Teams allows us to be anywhere all at once, right? I think one interesting trend I've also noticed is some of the best companies in the city are distributed in nature. Like ourselves, we are based both in New York and San Francisco. I see a lot of companies who were born out of SF and then moved to NYC because again, they want to be closer to their enterprise customers. So I think majority of the companies that you're going to see in the next five to 10 years are really going to be A, distributed. They're going to have presence in both the ecosystems because you just can't leave out New York City, right? It's too critical-

>> The speed is everything, Mohak. I mean, you got to move fast and have the complete frictionless comms build out customer value. Great stuff. Thanks for coming on and have a great event in New York in May, the AI Agent Conference. Thanks for coming on and taking the time. I know you're super busy. Certainly, let's do it again.

Mohak Sharma

>> Yep. Thank you for having me.

>> All right. Mohak Sharma is co-founder and CEO of HoneyHive, a New York-based company. Again, New York, Silicon Valley are connected, Wall Street and Silicon Valley through theCube and the NYSE Wired community. I'm John Furrier, your host of theCUBE. Thanks for watching.