What Happens When AI Agents Take on Your Enterprise Data with Rajiv Shah of OpenHands

Autonomous AI agents are rewriting what engineering teams can accomplish, but the gap between a working demo and lasting enterprise value is wider than most expect. In this episode, Rajiv Shah, Agentic AI Engineer of OpenHands, joins us to explore outer-loop coding agents, enterprise governance challenges and the persistent human role in making sense of organizational data. 

‍

Key Takeaways:

‍

00:00 Introduction.

02:41 Outer-loop coding agents execute autonomously for hours without hand-holding, from vulnerability scans to legacy code modernization.

05:25 AI agents with broad enterprise permissions create governance, security and data risks that demand sandboxed infrastructure.

11:09 Open-source AI platforms offer enterprises transparency, auditability and model flexibility to avoid vendor lock-in.

15:17 Moving from a demo to enterprise value requires change management, leadership buy-in and workflow integration, not just working code.

22:39 Data agents still need a semantic layer because human definitions of revenue, customers and churn vary by business and by quarter.

24:31 AI agents advance fastest in verifiable, deterministic domains and will continue reshaping any field that works in bits and bytes.

‍

Resources Mentioned:

‍

OpenHands website:

https://www.openhands.dev/

‍

Rajiv Shah on LinkedIn:

https://www.linkedin.com/in/rajistics/

‍

OpenHands on LinkedIn:

https://www.linkedin.com/company/openhands-ai/

‍

‍

Rajiv: [00:00:00]

The ideas that we have in our heads aren't always translated in what's inside the data store, inside the database. And so we need a little bit of that connection there, because often it's just implicit or it's understood. Or these meanings could be changing. You know, what revenue is — that's something a company can change its definition of. Is it over the last quarter? Is it over the last year? So remembering that those things come into play, because yes, you can have your data analysis agent go off and do an analysis, but it could be very different than what your actual humans would do, because they have different definitions of what's going on.

# Intro

Anthony: [00:01:00]

Welcome back to the Data Masters podcast. If you've been tracking the explosive growth of AI coding agents, you're gonna love today's episode. Joining us is Rajiv Shah, an agentic AI engineer currently on the founding go-to-market team at OpenHands, the leading open-source platform for cloud coding agents.

Raj's path to AI is fascinating. He started out in the military, where he earned the nickname "the Why Guy" for his relentless curiosity, which led him to a PhD and almost two decades of teaching data science and a career driving adoption at some of the biggest names in the data space, including Hugging Face, Snowflake, and Contextual.

Rajiv thrives in that messy zero-to-one phase of product building. He's passionate about outcome engineering and helping teams bridge the gap between complex technical tools and actual business value. So Rajiv, great to have you on the show. Welcome.

Rajiv: [00:02:00]

Excited to be here.

Anthony:

Awesome. So you recently joined OpenHands, and I think it's even fair to say OpenHands recently was launched. And as part of that founding, you've noted this shift from local coding agents and local coding assistance to what you call the outer loop. A lot of people listening are quite familiar with standard AI co-pilots that they run in their IDE. So in your view, what is this outer loop, what are you doing at OpenHands, and why do you think it's a big deal?

Rajiv:

It's an exciting time in AI, and I think part of the reason I went to OpenHands is just what we've seen in AI. A couple of years ago, coding agents were — if you remember — tab complete. They just did a tiny bit. But what we've seen, especially in just the last three or four months, is that these coding agents' capabilities have dramatically improved. Now they can put together a plan, do a series of tasks, and do this autonomously.

So what this means is we no longer have to use coding agents in that tab-complete mode where I always have to babysit it and only let it do one or two things at a time because otherwise it can go off the rails. Now I can really step back and just let it go — could be five minutes, could be half an hour, could be 30 hours, could be 30 days — and just let it code away toward some task. This ability to work autonomously now creates a lot more possibilities for use cases, whether in coding, data, or other areas.

Now I can set up jobs that run automatically — whether I want to scan my repos and look for vulnerabilities and automatically generate fixes, or perhaps I have legacy Spark code that I want to modernize and upgrade. I can have a background agent go out and test to see which of those repos should be upgraded to a newer version of Spark and automatically run tests to calculate what the efficiency gain is from moving the code. So that outer loop is this ability to do all these autonomous tasks, and I'm really excited about it.

Anthony: [00:04:00]

So this feels and sounds very similar to something many people have probably heard of and may have in fact installed and used — this tool, Open Claw. And I think it's fair to say there's been a sense of viral usage of it. Many people call it a privacy nightmare because you sort of give it unrestricted access to your machine. There are even some fun blog posts of people writing that it's done horrible things like deleted all of their documents. And I'm confident, talking to enterprise companies, that they would shudder to think of people installing Open Claw and setting it loose on, God forbid, their source code or even their data. So talk to me a little bit about what you're thinking about here at OpenHands versus Open Claw.

Rajiv: [00:05:00]

Yeah, and it reminds me of that phrase: with great power comes great responsibility. Now that these coding agents are capable of doing so many tasks and I can integrate them with many tools — Zoom, Salesforce, Snowflake — and have them seamlessly work and move data back and forth from all these things, go through emails and text messages, that opens up, especially for enterprises, a whole host of security, privacy, and governance issues. If you suddenly have agents making their own decisions to access data using some user's permissions to go in and get that data and do things — we've seen nightmare scenarios where companies have had their entire data, including backups, deleted because the agent thought, "You know what? We need to clean up and start over."

So there is a huge downside to trusting these agents to have access to all the data and permissions to do all kinds of commands. And this is where, especially at the enterprise level, a governance layer comes in — there are sandboxes, there are controls around that to harness this capability while still not putting ourselves too far at risk.

Anthony: [00:06:00]

So as a practical matter, how do you implement some of this? Because I think one of the things people like about these autonomous agents is they do have their permissions and they are able to do work. It's a little bit like giving an intern access to your email or letting them borrow your credentials. Maybe not a perfect analogy, but it certainly has this sense of multiplication — like me, but five of me. And that's exciting.

Rajiv: [00:07:00]

Yeah, and obviously there's gonna be a tension here, because we all want to get things done and if we can use these agents to get things done — but on the other hand, are you as an individual going to download huge amounts of data and move it to another system? An agent might do exactly that. So yes, we want to use your permissions and make life easier for you, but we still want some type of governance and controls over that to make sure they're not moving data they're not supposed to, not doing actions they shouldn't be doing, not downloading commands or running unauthorized software. Because maybe for the agent it could be a shortcut, since they don't understand all the context or all the rules an enterprise may have in place.

Absolutely, it's gonna be a tension. We see it nowadays with tools like Claude Cowork and Claude Code, which are really based on a laptop user working entirely on their laptop — versus many enterprises that don't want that much going on on laptops. They prefer things in a sandbox, in a place where they can control and know where the compute is being done. That's gonna be a source of tension as we go forward.

Anthony: [00:08:00]

That's a really interesting point. And it's something I hadn't thought about until you said it — there's a sense that a lot of these agents come from your local PC context, whether it's a Mac or a PC. And I suspect that's because that's how developers tend to work. Developers typically work on a very powerful PC. And I suppose it also maps nicely to this trend of people — the new hot thing to do is buy a Mac mini, install Open Claw, and let it loose. But again, it goes back to this idea that the core construct is my local machine, which is a local version of me. Are you doing something different with OpenHands that sort of breaks that paradigm?

Rajiv: [00:09:00]

Yeah, so I think there's always gonna be this tension — and we've seen it in IT — between centralization and decentralization. With OpenHands, we've spent a lot of time working on coding agents for several years. And one of the things we came to early on is that coding agents need a sandbox. They need a protected environment to work in. We don't want them running loose on laptops. So that's always been part of the architecture of OpenHands — to use secure sandboxes where we have network controls over what data is coming in and out and have that isolated environment to run things in. Because it's not only useful for security, but also useful for scaling, for compute, and for lots of other reasons. Maybe you don't have the fanciest MacBook laptop, but you need a lot of compute for your coding job — having it in a container sitting in the cloud gives you a lot more flexibility.

Anthony: [00:10:00]

Got it. So not only do I not have access to my local machine, I may actually have access to more compute and more capabilities because I don't have to actually manage it.

Rajiv:

Absolutely. And for a lot of the coding tasks — and you can think of this even for data tasks — it's not just one or two streams of data I want to work with. I might be working across an entire enterprise where I want to do a hundred or a thousand different things at the same time. That's gonna be way too much for one laptop to handle. So I think that's a key difference between personal use and a lot of the enterprise use cases, which take a lot more resources.

Anthony:

Got it. And I suppose that helps address one challenge people have had with coding agents — it does demand a fairly beefy PC, hence my Mac mini comment. And that just wouldn't be a good idea in the enterprise — you wouldn't want to turn that on for everybody. But in this context it makes a ton of sense. I'm also curious — you've been a big proponent of open-source, and this is an open-source project. Talk a little bit about how that fits into the strategy.

Rajiv: [00:11:00]

Yeah, and I'll give a little bit of history — I'm an old person. Part of how I started out in data science was because of open-source. When I first got into statistics, you had to go spend a thousand dollars to get SPSS from IBM if you wanted to run some social stats. It was open-source that cracked open the world of data science and AI and made it widely available to folks. So in my heart, I've always loved that approach.

And I've also seen it from the enterprise side. I remember the days when open-source was not looked upon favorably inside an enterprise — there were real concerns about security. And now we've seen enterprises flip-flop, where I talk to lots of companies that really have an open-source-first philosophy, because they want the ability to transparently look at the code, make sure there are no issues, and fix things themselves if there's a problem. And so I've seen that real benefit from working with enterprises, which is one of the reasons I wanted to choose an open-source company again. I spent many years at Hugging Face because I just know enterprises nowadays really value that.

I think the other piece when it comes to AI is being agnostic to the type of model choice. Everybody has favorites with things like Anthropic and OpenAI, but the model providers come and change — what's fashionable today is not gonna be fashionable a year from now. And realizing that you don't need to tie yourself into one model, because a lot of times we magnify the differences heavily between these model providers when for 90 to 95% of tasks you wouldn't know which model is running. So having a solution that's model-agnostic gives enterprises the flexibility to choose what fits best for that particular job — that's always been something I thought was very invaluable.

Anthony: [00:13:00]

Got it. So it feels like the open-source choice is also about trust. If I'm trusting this thing to run autonomous workloads, especially in the cloud, then having a sense of what the actual system is doing and being able to evaluate it is an important core component.

Rajiv:

Yeah. And it's become so much easier nowadays, because now we have all these great coding agents. When somebody hands you some open-source code, you can point your coding agents to it and quickly understand exactly how it works — what the interfaces are, how you would modify it or change it — where some traditional software was much more opaque and you didn't have that ability.

# Sponsor Plug

Anthony: [00:14:00]

Let's shift a little bit into thinking about the value that users get out of these coding agents, and particularly in the context of data projects. I know you've spent a lot of energy thinking about how to help companies and people take advantage of these technologies and actually get business value out the back, as opposed to yet another AI project that didn't quite go the way people expected. Maybe talk a little bit about — and I think you call it outcome engineering — how do you think about helping people get value out of these tools, especially in the context of data projects?

Rajiv: [00:15:00]

Yeah, I think there's one element of how we use these tools. But one of the things I'd also like to focus on is the larger piece of how we make these tools work inside of an enterprise. And I'll tell you, I'm the first one guilty of this — I will put out on social media shiny demos: "Hey, look at what I built!" And I remember working inside of enterprises, showing up on a Monday morning with some weekend project: "Hey, look what we did. We integrated these pieces together." And it becomes very easy to set up and create those types of demos.

But to really create value inside an enterprise, you have to go much further beyond that demo — you can't get stuck at the pilot project stage. You have to think about the fact that in an enterprise, there are a lot of people working there. Whatever you build — yes, you are the power user, it works for you — but you've got to make it work across everybody. It has to fit into their workflows. You have to have the change management and the education and the incentives for people to want to learn a new workflow. You need to have it plugged in so that what you built is actually valuable to the business in terms of what the business cares about. You want your leadership bought in, you want a VP on your project so that it can move forward. So when we talk about value, that's an important piece — how you actually get that cool piece of technology all the way so it's valuable inside an enterprise.

Anthony: [00:16:00]

Okay. So talk about that, because I think a lot of people will have had the experience of being able to do much more because they have an AI assistant or a coding agent helping them — hacking something together, for lack of a better term. Although maybe that's actually a bad way to put it, because there's no longer hacking things together. You can actually write production code without even having to understand in a surface way what the code is doing. You can achieve fairly sophisticated outcomes without a real understanding of the details. So this idea that the weekend project can become a production application — we're much closer to that today than maybe we've ever been.

Rajiv: [00:17:00]

So I think we're closer to that on the technical software engineering part. But as a process for getting that inside enterprises and connecting the training and leadership and all of those pieces — we're not. And if we look at the engineering side, this has changed a lot over the last year. Six months ago, if somebody showed up on a weekend saying "I built this using the latest AI models," a software engineer might shrug and say yes, it kind of works, but there are concerns about reliability and edge cases and how it would break. Now, in the last six months, the quality of the code is much closer to what you'd expect from developers. If you're working in popular languages doing popular applications, these coding agents really know how to work well. They can develop production-level quality code now. There's still some quibbling among software engineers about coding best practices and reliability, but we can all agree it's much closer now.

But I think this is where the trap is — yes, you've built the code, but to actually create value inside an enterprise, you need much more than that code to change the business processes and create value.

Anthony: [00:18:00]

Got it. Okay. I suppose it's an open question as to when we expect coding agents to also help drive business adoption — sending emails, asking people to contribute. And I suppose we see a bit of this already: asking an agent to write the documentation for the code you've written, or the user adoption manual, or the use case reports, or even the PowerPoint presentation describing it. These are things we could imagine doing today.

Rajiv: [00:19:00]

Absolutely. And again, I think we have to think through the implications of this. We've moved to a point now where these coding agents are so good at writing documentation, so good at writing PowerPoint decks, so good at writing emails — that the human side of us can get kind of overwhelmed or annoyed by how much content is coming through. Some of that content is obviously for the agents themselves to help understand what to do. And this is where I still think thinking about the value inside the business organization is important for how it moves ahead — what are you actually trying to do? Sell more widgets, lower costs, that kind of thing.

Anthony:

Wouldn't you say — I think we've maybe mixed up two ideas here. Writing code for the purpose of building an application — let's just grant you that for a second.

Rajiv:

Yeah.

Anthony: [00:20:00]

For data projects, the bar is probably lower. If I'm writing a bunch of Python code to marry two different datasets and come up with some insight about our gross margins, for one thing it may not need to be production code — it may just need to answer the question. And secondarily, the ability to ask a coding agent to run a statistical test, interpret the results, and write it up in a way that an executive can understand — that's probably a pretty easy task for agents these days.

Rajiv: [00:21:00]

Yeah, it's a really interesting terrain around data analysis, because coding agents definitely have the ability to manipulate data and apply the tools and statistical knowledge to do these tasks. But where the rubber meets the road is often the data isn't pristine. There are still a lot of interpretations. You have eight different versions of the date column, and so there's still that human partnership that has to come into play to help explain how the data is arranged and how we actually calculate revenue or subscribers. Because some of that stuff might not be documented. So absolutely it's gonna accelerate that analysis on the back-end piece, but in the companies I see, it's still pretty messy in terms of how their data is arranged and what their business objectives are — there's still a heavy human element required to figure things out and squeeze out all of this value.

Anthony: [00:22:00]

Well, that's a great point. And maybe to play it back to you — it might be fair to say that if everybody had their data perfectly organized and linked together, there was this dream from, I don't know, the eighties or nineties, that the end game was everybody having all their data in SAP. If we had achieved that dream, then maybe we would be in a place today where we could just say "run the margin analysis" and go. But your point is that there's still a very big missing piece — this contextual layer for the data that brings it together, organizes it, cleans it, et cetera. Is that fair? I don't want to put words in your mouth.

Rajiv:

Absolutely. And it's just that contextual or semantic layer, because the ideas that we have in our heads aren't always translated in what's inside the data store, inside the database. And so we need a little bit of that connection there, because often it's just implicit or understood — or these meanings could be changing. You know, what revenue is, is something a company can change its definition of. Is it over the last quarter? Is it over the last year? So remembering that those things come into play — because yes, you can have your data analysis agent go off and do an analysis, but it could be very different than what your actual humans would do, because they have different definitions of what's going on.

Anthony: [00:23:00]

Yeah. That's right. And also the interpretation of what a customer is, or a customer who churns, or a customer who upsells — those things may have good generic definitions, but they have very different meanings in your business.

Rajiv:

Exactly. And you know, this is the same revolution we're seeing in coding and data science — these agents are now really good at harnessing these tools, but we still need a little bit of human oversight, especially in messier areas, to help guide them on what makes sense to work on and what doesn't.

Anthony: [00:24:00]

Got it. So casting your eye forward — and it does feel like this world is changing faster and faster, almost hard to keep up — asking for a five-year prediction is almost certainly useless. So thinking six to twelve months out, what do you see shifting and changing? What do you expect people will be doing differently in the medium term?

Rajiv:

Yeah, so thinking back to why AI agents are so good at, for example, software and math — it's because they're very deterministic tasks. There's a kind of right or wrong answer. With coding, there are lots of checks in place partway through where you can verify you're on the right track. We have tools like compilers —

Anthony:

Right. Things can compile, they can type-check.

Rajiv: [00:25:00]

Exactly. Because of that, it was an area we were able to rapidly advance. Now the coding agents are as good as the best humans — better than 99% of coders out there. So now the question is: what are the next areas? I think looking at how coding agents have been successful in these verifiable tasks — where there's a clear right or wrong answer and tools to identify whether you're on the right track — those are gonna be the areas most upended next.

I think the other piece we see is the rise of skills — being able to codify specific instructions to agents to do tasks. More and more we're gonna pull that out of humans to train these agents to do more and more. The growth of these agents has been amazing over the last two years and I really don't see any reason these models are going to slow down in their abilities and capabilities. Anything that works in the digital domain — working with bits and bytes that can be manipulated and read by a computer — these agents are going to be working on all of that. Whether they're creating images, doing data analysis, or writing legal briefs, this is going to be a substantial change.

As humans, we've had technological change before and we can adapt to it. Sometimes we overthink the impact in the short term, but over the long term this is going to be a revolutionary impact as these models can harness so much more digital information and manipulate it much better and quicker than humans can. So it's gonna have some long-lasting effects.

Anthony: [00:27:00]

Yeah. Maybe to put a positive spin on that — there have been a lot of challenges in the enterprise, and we talked about one around cleaning and contextualizing your data, that maybe were unsolvable until we had access to these kinds of technologies. Certainly, we will see more software written in the next six to twelve months than we saw written in probably the last six to twelve years. If that pace increases — is that a fair, more positive framing?

Rajiv:

Yeah. And I don't mean to be negative. I think there is gonna be a dramatic impact — we're not exactly sure what it's gonna do. But the history of technology has shown it's made life better for humans overall. There's gonna be some impacts along the way, but I think it's an exciting time to be in, where these machines are being able to free us from a lot of tasks so we can work on higher-order pieces. So yes, it's an exciting area with big changes ahead.

Anthony: [00:28:00]

Awesome. Well, Rajiv, a pleasure. Thank you for sharing these insights. I think there's a lot here for people to take away and hopefully use and make practical in their day-to-day.

Rajiv:

Thank you for having me. This has been fun.

# Outro

FLAGGED FOR REVIEW

Unverified Proper Nouns

- "Open Claw" — [00:04:00], [00:05:00], [00:08:00] — Anthony refers to a tool called "Open Claw" (also spelled "open claw" and "open cloud" once in the raw export). This may be "OpenCLAW," "Open Claw," or a misrecognition of a different tool name entirely. Please verify the correct name and spelling before publishing.

- "Rajiv Shah" vs. "Rajiv" / "Raj" — Intake spells name as "Rajiv Shah." Anthony addresses him as "Raj" at [00:01:00] (used his nickname) and as "Rajeev" at [00:28:00] (likely Descript mishear of "Rajiv"). Corrected to "Rajiv" throughout the cleaned transcript. Please confirm preferred on-air name.

- "the Why Guy" — [00:01:00] — Anthony's intro describes Rajiv's military nickname as "the Y guy." Interpreted as "the Why Guy" (a curiosity-driven nickname). Please confirm spelling.

- "SPSS from IBM" — [00:11:00] — Referenced as a paid stats software. Verified as a real product; spelling confirmed.

- "Claude Cowork" — [00:07:00] — Rajiv references "Claude Cowork and Claude Code" as tools. "Claude Code" is a known Anthropic product. "Claude Cowork" may be correct or may be a Descript mishear of another tool name. Please verify.

Structural Notes

- The Quote section at [00:00:00] is a standalone pull-quote in the Descript export. The same passage recurs verbatim later in the conversation at [00:22:00]–[00:23:00]. Both instances are preserved.

- Anthony's line at [00:11:00] — "And I suppose it also maps nicely to this trend of people. Uh, I think the, the new hot thing to do is buy a Mac mini install open claw" — appears to be a single sentence split by Descript. Rejoined in the cleaned transcript.

- Anthony's line at [00:18:00] — "last question, and this is the one you always have to have a heads-up on ahead of" — trails off in the raw export mid-sentence. In context, this appears to be a throwaway aside before he pivots to asking about books. Removed the trailing fragment; cleaned so the question flows naturally. Editor may wish to review audio.

- Timecodes appear only at section-level in the Descript export and do not appear within most speaker turns. Preserved exactly as they appear in the original.

Inaudible Sections

No [inaudible] tags found in original transcript.