Swapnil Bhartiya: You might think that managing one agent is hard enough, but the reality is most companies there is now an explosion of AI agents, each running independently, often without any oversight or coordination. That is a recipe for disaster, confusion, risk and of course, wasted efforts. Now OpenHat is tackling this head on with their new agentic control plane designed to bring all all those agents under a single pain, a blast. And today we have with us Robert Brun, co founder and CEO of Open Hands, to talk about how this changes the game. First of all, Robert, it’s great to have you on the show.

Robert Brennan: Yeah, it’s great to be here. I appreciate you having me.

Swapnil Bhartiya: It’s my pleasure. Of course we have hosted Open Hands before, so our audience, they do know about the company, but since you are here. So I would love to know a bit about what specific challenge you saw in the space that which led to the creation of this company. Let’s start there first.

Robert Brennan: Yeah, so we’re really focused on helping folks graduate from prompting with agents to building processes with agents. I would say most developers today are comfortable kind of like piloting an agent on their laptop pair programming, but that’s not super scalable. It is a good lift in productivity for that day to day ad hoc feature development. It typically gives like a 20 to 30% productivity boost. But there’s a whole bunch of productivity gains to be had if you can start to run agents automatically. Agents that run in the cloud in response to events coming out of third party systems or running on a schedule and don’t need to take up space in the developer’s brain. Right. For instance, you can have an agent that every Monday morning updates all of your dependencies or scans your repo for vulnerabilities. You could have an agent that listens to Datadog and every time an error occurs, it immediately investigates the error and opens up a pull request to fix it. There’s a lot of just like toil and rote work that goes into keeping a code base healthy. And it no longer makes sense to have engineers dedicating some portion of their attention to those problems. They’re problems that can be easily solved by productionized agents.

Swapnil Bhartiya: Excellent. Thank you. And that’s actually we also focus here to talk about not the hype, but actually how AI is going to production. Now as I was talking about, the reality is that organization, Even individuals or SMEs, they don’t run just one agent. And also it becomes kind of tempting and actually in some cases actually better to have a specialized agent, even if it doesn’t matter which model you’re using for this specific task? Because agents actually, latest models actually are more troubling than the past models because they’re trying to increase efficiency, which is actually crippling those models. Can you talk about, based on the insights you folks have about the industry, how many agents are most company actually running at any given time on average?

Robert Brennan: Yeah, it’s a good question. So we focus specifically on software development agents. So agents that are modifying code or you know, participating in the software development life cycle in some way. You know, I would say most companies right now they’re not running any agents in production. They are still kind of experimenting, you know, prompting Claude code on their laptops and that, you know, that’s, that’s a reasonable place to start. I would say the most forward leaning companies are running, you know, probably hundreds of agentic sessions per day, you know, thousands depending on the scale of the company. Typically they start with something like code review, something that’s fairly easy, that’s not going to make any changes to the code base. Right. It’s just basically every time a pull request gets opened, an agent is immediately triggered to go and review the code that’s been opened up here and decide, is this a change that’s worth merging? Are there security problems here? Are there QA issues here? That’s a huge lift for the team. Code review is a huge bottleneck right now, but that’s kind of like the first pass at putting an agent into production that’s going to trigger automatically rather than be piloted by an individual engineer. But very quickly folks will start to expand beyond that one use case to a host of different use cases. Right. Dependency management is a big one. Security vulnerabilities. Solving security vulnerabilities is a big one. Doing pen testing is another big one. Managing things like release notes, there’s a whole bunch of like toil that goes into the code base. So I would say typically folks start with like one code review agent that’s going to run, you know, dozens of times a day. Every time a pull request gets open, maybe thousands of times per day. If you’re a very large company and they expand to other use cases and get more and more agents going.

Swapnil Bhartiya: What are some of the biggest challenges or headaches team face when they have to manage all these agents at scale?

Robert Brennan: Yeah, it’s really tricky. I think a lot of teams start by kind of hacking something together with like GitHub Actions and Cloud code. You know, they’ll kind of piece together the tools that they’re used to in order to get something working and that’s great for like one use case. But it doesn’t, it doesn’t scale very well. Right. What you really need is some centralized visibility into what agents are running. You know, how are they performing, are they, are they being successful or not? How much are they spending, what’s the security profile of these agents? And so the, the highest performing teams are find, are building or buying a central platform in order to drive these cloud based agents to work on, you know, how do you provision a sandbox for the agent, how do you measure its spend, how do you measure its success rate, etc. And that’s really where we focus on with our agent control plan at Open Hands is really helping folks understand, okay, you’ve got a code review agent. We can see that 85% of its comments are getting accepted and merged into the PR. So it’s a fairly successful agent. It spent $2,000 in the last month. That’s going up 15% month over month. Like we can help you understand the ROI of each individual agent so you can know where are you actually getting value and where are you just burning tokens.

Swapnil Bhartiya: Now let’s talk about the agent control plane. How does it help teams keep tabs on all those agents? One more thing. I think a few days ago Anthropic also released their agent manager. I think how is it different for. Because organizations are realizing that agents are becoming a problem. So first of all let’s talk about your control plane and then let’s compare with what things like Anthropic are offering there.

Robert Brennan: So there’s really two halves to the control plane, right? It’s visibility and it’s control, right. So number one, you just want to be able to know where are all these agents running and what are they doing? How much are they costing? How is that changing over time? When you’re patching together systems and every team’s patching together their own solution for running agents in production, you don’t really have that visibility. You don’t really understand what agents are out there, what are they doing, what’s the security profile, et cetera. So getting that visibility I think is the first piece. And then it’s about control. It’s about understanding, okay, we see this agent is running, it’s only got this level of success rate, right? This dependency updater agent is only getting 10% of its PRs merged. It’s burning thousands of dollars worth of tokens. We need to shut this one down or iterate on it, make it better getting some control over like what models are being used, some control over like what the sandbox these agents are being put in is like what secrets do they have access to, what tools do they have access to? Being able to create separate environments for like an agent that needs, you know, production level access versus an agent that just needs to do code review. That’s really what the control plan is about, is giving, giving the management team and the, and the executive leadership a sense for what’s happening here. Right. You give the CFO some visibility into like what are these agents costing, which ones are driving ROI? You know, you give the CISO some visibility into like what is the security profile here? Which of these agents are actually dangerous, how are we running them, what models are we running with them? You know, it’s, it’s, it’s really about giving the whole, the whole team what they need to make intelligent decisions about a technology that is very powerful, very dangerous and very expensive.

Swapnil Bhartiya: And how is different from what Anthropic is doing with their cloud managed agents? You know, it’s kind of like building harness around agents.

Robert Brennan: Yeah. So managed managed agents is really about that like orchestration layer, getting the agents firing on a schedule or in response to events coming out of the cloud. I would say one of the big downsides with going with Anthropic’s platform, for instance, is that you’re really getting stuck into a vertically integrated solution. And I think a lot of companies are realizing that they need to be able to switch between models. Right. They don’t want to be just an Anthropic shop or just a Gemini shop or just a codec shop. The models are fairly commoditized at this point. There’s some differences in quality and speed and cost, but they’re constantly leapfrogging each other. And very frequently, you know, folks will build around the Anthropic ecosystem and then OpenAI will come out with something and they’ll be like, oh, we need to like jump ship and rewrite all this stuff so that we can use the latest OpenAI model and then they’ll want to swing back to Anthropic later. So going with a vendor who is model agnostic I think is super important. And it also gives you some, you know, if your team is naturally going to be using many different agents, right. Some people are going to want to use codecs, some people are going to want to use Claude code, some people are going to want to use the openhands harness. Anthropic is not going to want to wrangle all that for you, right. You’re going to want to have a third party provider that can pull all of that into one place. And Open Hands does work with every harness out there.

Swapnil Bhartiya: Can you talk about some of the core features and let’s look at it from the perspective of teams. Core features of control plane that actually makes life easier for teams who are dealing with multiple agents.

Robert Brennan: Yeah, I think the biggest thing on people’s minds right now is budgeting. You know, people are starting to really see these bills pour in from their, their token spend. And so one of the things you get, you know, immediately out of the box is the ability to set a budget at an organization level, set a budget at a individual user level, at a team level, and I think most importantly at like a use case level or like a, you know, a per agent level. Right. You could say we want to spend this much on code review this much on dependency updates, et cetera. I think that’s the first thing that people look to is like, that’s the, that’s the problem that’s most top of mind is like how much am I spending? Where is that spend going? Who’s, who’s causing that spend? I think another, another big piece is increasingly agents aren’t just being owned and piloted by individual engineers, they’re being owned and piloted by the team. Right. So a code review agent for instance, doesn’t make sense for that to belong to me personally or for me to have a separate code review agent from one of my coworkers. Right. We want to be using the same code review agent on every single one of our repos so that, you know, we’re getting, we’re getting similar results and whatever learnings that are happening in that process are spread across the entire team. Right. So now that agent, we need to be able to control who has access to see what the agent’s doing, to control its behavior, to decide what model it’s using to decide what its budget is. And so you get some role based access control that allows, you know, say the, the team manager to set the model and the budget. Maybe anybody on the team can change the prompts that, that drive that agent. So some of that, that access control now that agents are really owned by the team rather than individuals is really important. And then I think the biggest and most exciting piece is really this ability to measure ROI, this ability to take an agent and feed a success signal back into the system to understand how good is this agent. Is it actually delivering value to us or is it burning tokens? Right. So on our code review agent, for instance, we feed a signal back into the system about how often the comments that the code review agent made are actually getting resolved. Or we could feedback information about additional comments that were made by a human in a second pass review that maybe the agent missed. Right. And so we can understand how accurate is this agent, what is the precision and recall of its, of its comments? And we can actually see, because we have this data, we can see our accuracy rate going up over time for our code review agent because we’re constantly, we’re constantly improving in response to those signals. So you get a feedback loop with the agent control plane that allows you to improve your agents over time.

Swapnil Bhartiya: As important as it is to manage agents, there are certain things which are more about how AI works these days. Depending on the industry, or even if you’re not in the compliance or regulated industry, governance has become a big challenge. Security, of course, is becoming a big challenge. How much access you should give or not give is a big challenge. Of course, hallucination is a big challenge. There’s only so much you can do through the agent control plane. But how do you ensure at least some of the critical pieces, let’s say security governance, when so many agents are operating independently?

Robert Brennan: Yeah, yeah, no, it’s a great question. I think this is one of the biggest boosts of the agent control plane. And I think security is becoming a bigger story now. And I think probably over the next year or so it’s going to become a very critical story. The nice thing about running agents in the cloud that are owned by the team, rather than running agents on everybody’s individual laptop, is that everything in the cloud is like naturally sandboxed and adhering to principles of least privilege. Right. When you’re running an agent on your laptop, the agent has access to everything you have access to, right? It has access to all your files, it has access to, you know, if you’re connected to a production database, the agent can like muck around in there. If you’re connected to a Kubernetes cluster, the agent can see that. So when you run an agent in the cloud, it gets its own sandbox, it gets its own environment. And you can specifically decide what secrets should this agent have access to, what system, should it have access to, what MCP servers, what tool calls should it be allowed to make? And you can set that on an agent by agent basis. So you can have one agent, say, for doing incident triage, that has access to your production datadog, but maybe just has read only access you could have another agent which is just doing code review, and then it doesn’t need access to the datadog at all. Right. It just needs access to the code locally. So maybe it just has access to GitHub or something like that. So you can decide you can basically lower your blast radius by cutting down what each agent has access to. Right. We call this the principle of least privilege. It’s been a principle in software development and security for a very long time, and I think we’ve kind of forgotten it when it comes to agents that are starting to remember how important it is. So, yeah, a lot that can be done on the security side. I think there’s also some really important things around, like auditing, being able to retroactively look at what agents have done, see their traces, see their trajectories. And again, that’s something that you really get from an agent control plane because now you have a centralized place where all those trajectories are being stored. You can go back and look, if something does happen, you can decide, okay, what, what went wrong here? What did the agent do? Did it hallucinate? Did we give it too much access, et cetera?

Swapnil Bhartiya: How should organization look at AI and agents either differently or the same way we look at software, that not as much about technology, it’s more about having right practices, right processes, right culture in place.

Robert Brennan: Yeah, I 100% agree with you. I think right now a lot of companies are in the experimentation phase. They’re messing around. Their individual engineers are putting together little proof of concepts. They’re wiring things together on their own laptop. Everybody has direct access to agents. I think that’s good. I think we’re still trying to figure out where we can get the most value from agents. But there needs to be a very clear process for taking something from the experimental phase runs. On my laptop, I was able to do this in one repo to something in production where I can now take this agent and scale it across the entire organization, get it running automatically instead of, you know, running just on my laptop when, while I’m driving it. I think that having that kind of process of how do you, how do you take an agent from experiment to production? Is, is probably the most important process. I do think it’s, it’s a little bit of a mistake to think about agents as co workers or as employees. I think there’s, you know, there’s a, there’s a famous quote from IBM that a computer can never be held accountable. And I think that’s, that’s still very true. Right. You. Even if we’re getting more and more done with agents, I do think there needs to be a human being who is ultimately responsible for their behavior so that when something does go wrong, it’s not the CEO like trying to, you know, prompt his way back into a better state. You know, there’s somebody who’s a great systems level thinker who understands, you know, how these agents have been set up, what they’ve been instructed to do, how they behave, who is, you know, looking at that system and figuring out how am I going to rearchitect this so this problem doesn’t happen again. I think there’s always going to be a place for those, you know, that, that engineering mindset, whether they’re working with source code or they’re working with agents. You know, maybe, maybe we’ll go from software engineers to robot psychologists, but I think there’s going to be a need for that, that engineering brain.

Swapnil Bhartiya: Can you talk about what does the adoption look like for engineering managers or dev team today and how, of course, it has been out for a while and before that you may have run some beta tests with your users as well. What was that? Their feedback. How, how did their day to day operation headache change where they’re like, no, this is actually solving one of the biggest problems that we are tackling there. So I would like to understand this two point question. One is the adoption. The second is what is their reaction, their feedback?

Robert Brennan: Yeah, I would say the, there’s, there’s a little bit of a tension between the platform teams who we typically work with on agent control plane who are trying to like put in some process and you know, figure out how are we going to do this securely, how are we going to measure budget, things like that. And the individual engineers on the ground who just want to, you know, they want a blank check to run all the agents they want and do whatever they want. Right. I would say we get a lot of great feedback from the platform team who’s like, hey, like now I, now I have eyes on this, on this problem I’ve started. I have a process to go from prompting to production. You know, I have a platform here that I can now start to build on and really start to have a clear sense of like this is how we run agents at our company. So the platform team tends to love it. I think because there are increased guardrails in place, we do sometimes hear pushback from individual engineers on the ground who feel like something’s being taken away if they no longer have control over everything. That’s running on their laptop. You know, if they’re being pushed now to run agents in the cloud, I would say that is tempered by excitement that they no longer have to pilot every single agent that they’re running. Right? They’re able to take an agent that they’ve been like manually typing into saying do the thing over and over and over again, do a code review over and over and over again on their laptop. And now that’s just happening automatically in the cloud. And that does feel like magic. And I think that is enough to overcome the fact that that code review agent now has some more process some more guardrails in place around it.

Swapnil Bhartiya: When it comes to general software development, developers are the king. They pick their tools, they pick their frameworks and they make the decision. The right tool when it comes to AI is it trickled down. There’s a bottom up that organizations because you have to sign mega million dollar contracts, right? It’s not that, hey, you can just go sign up chat GPT because also about you have to put a lot of, you know, data sensitive data and you cannot just load everything in somebody’s cloud because even through API all the data is going through that. So when it comes to AI, who is calling the shots? Is it top management or is it still developers who are in or what are you seeing in the market? What kind of struggle do you see there?

Robert Brennan: Yeah, so what we see is there’s typically two, two workflows, right? There’s for the devs who are, you know, pretty much every developers is still running agents on their laptop. And I think that’s, that’s not going to go away. Right. For certain tasks where you don’t have a clear definition of done, you’re still kind of exploring the problem space. It makes sense to have a quick feedback loop running on your laptop with an agent. And there we see companies typically offering a menu of options to their developers and saying hey, okay, we’ve got cursor licenses, we’ve got windsurf licenses, we’ve got Claude code licenses. Like pick the thing that works best for you. You’re working on your laptop, this is your environment. The same way you’ve picked your IDE for years, you’ve picked your own local tooling. You can keep doing that with AI, but the company will typically have a set of licenses that you can pick from. But then in the cloud when they’re productionizing agents, that’s when they center on a single system like Open Hands, they pick a single platform and say hey, this is how we’re going to take an agent that’s being prompted on your laptop and productionize it so that it’s running at scale automatically. And it makes sense. It’s kind of similar to how every developer could choose their own IDE, but the company would pick GitHub or GitLab or BitBucket as their Git provider. It doesn’t make sense if I’m pushing my code to GitHub and you’re pushing your code to GitLab. We need to collaborate at that cloud layer, at that collaboration layer. And so that’s where companies are really centering on a single platform to run their agent.

Swapnil Bhartiya: Do you see a future where there are going to be more AI agents than employees inside most organizations? Because this is what we’re hearing from a lot of companies. Lay off all the people, get two or three employees, high profile employees and just a fleet of agents. So the job is to manage those agents versus hiring a lot of, you know, mid tier or low tier employees. Where do you see things are going?

Robert Brennan: Yeah, I do think there’s going to be a lot of agents doing a lot of work, but I don’t see them replacing humans. Agents are quite good at kind of the rote work. When there’s a really clear specification of this is what we want to build and this is how we’re going to build it. The agent can go and execute that very, very well. But you still need, basically there’s two categories of human beings that are going to need for, as far as I can see into the future, going to need to be a part of the process and part of a company. One is on the product side, you need people with clear vision for, you know, what are we building and what do we want to build, how are we going to build it? You know, agents are great at executing on specific instructions, but if you try and brainstorm with them, they come up with all sorts of wild ideas. Like they’re, they’re bad product managers, they’re good, they’re good aids for a product manager, but they’re, they’re, you wouldn’t want to just like have an agent looking at all your user feedback and deciding what to do next. You’ll end up with a Frankenstein sign of a product. So we’re going to need humans on the product side, but then we’re also going to need humans on the systems level, engineering side. You’re going to have hundreds, thousands of agents all interacting with each other. There’s going to be cost overruns, production is going to go down, things are going to tip over, weird things are going to happen and you’re going to need smart people who are looking at these systems and figuring out how are we going to re architect this so that it’s more stable, more secure, more reliable. So I, I think you’re going to need those still, those, those systems level thinkers, those, those good engineering minds, you know, they’re going to be one level of abstraction up, they’re going to be getting a lot more done. But I like to say, you know, we’re, we’re going to be, we’re going to have 10 times as much software in the world in a few years. Right? We’re going to need more software engineers to manage all that. Right. We’re going to be able to get a lot more done. But the, we’re still going to need a lot of smart people thinking about what’s going on here.

Swapnil Bhartiya: Talk a bit about what advice do you have for companies all the way from board level, C level executives to teams who are just starting not their AI journey, but their agent journey, it’s

Robert Brennan: still very important to experiment and to enable individual people at the company to experiment. We see a very wide range of like AI skill sets. Some people are very eager to experiment and adopt and like they, the, the, you know, running with agents just kind of makes sense to them and they figure out how to use it very productively. Other people are a little bit more resistant, a little bit more stuck in their, in their existing processes. So I think giving, giving everybody within the organization the, the autonomy to experiment is, is really important. But I think there’s also a really strong need to figure out once you do identify a great use case for agents, how do you scale that across the organization? How do you actually productionize that? And that’s where an agent control plane really comes to play, is that it gives you a process for taking something that one of your interns figured out, a way to do X, Y and Z with agents. How do we get that off of that intern’s laptop and into prod? I think that’s a really important process to have in place.

Swapnil Bhartiya: Robert, thank you so much for joining us and sharing your vision for agent management. And of course those who are watching, please check out Open Hands to learn more about their agentic control plan. And once again, Robert, I would love to have you back on the show.

Robert Brennan: Thank you, that’d be great.