This week on Dev Interrupted, Anush Elangovan, VP of AI Software at AMD, returns to unpack the rapid shift toward an agentic software development lifecycle. Anush introduces the concept of "Agentic IO," a workflow where engineers focus strictly on high-level goals while AI handles the complex implementation. The conversation also highlights the expanding productivity wingspan of modern developers, the power of local open source models, and why speed remains the ultimate competitive moat.
Show Notes
AMD ROCm: Learn more about AMD's open-source software stack for AI at rocm.docs.amd.com and on GitHub.
AMD Advancing AI 2026: Register for AMD's flagship global AI event taking place July 22-23 in San Francisco at amd.com/advancing-ai
Follow Anush on LinkedIn: Anush Elangovan | AMD blog
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: I'm super excited to kick this off. So, you know, we're back, and joining me today is Anush Elangovan, VP of AI Software at AMD. And Anush, it's only been about six months or so since you were last on the show, but in the world of AI, that genuinely feels like a decade or two. And last time we talked about how speed is the moat. That was the thesis we walked away from with our conversation with you, and the importance of AMD's open source ROCm stack driving all of that. But since then, the industry has certainly evolved, and we're talking about a lot of things in orchestrating and delivering agents at scale. And hardware is still a fundamental part of that, but it's part of a bigger and more complex puzzle. So we're seeing a fundamental shift underneath with the folks that are utilizing and making the technology possible, a shift towards this agentic AI world. Like six months ago when we talked, we- people were barely [00:01:00] orchestrating, and now GitHub can barely stay online. So Anush, I'm so excited to dive into how the landscape is evolving, what your perspective on it is, and how AMD is rising to meet the occasion.
[00:01:13] Andrew Zigler: So Anush, welcome back.
[00:01:15] Anush Elangovan: Yeah. Thanks for having me, Andrew. This is exciting. Um, yeah. I mean, six months ago, I, I think, um, I, I probably thought, um, you know, agents were like, uh, prompts in a cron job, right? Like, it's just a, you know, periodic... You're just prompting something and, and it's giving you some information. but in the past three months, it's kind of like evolved.
[00:01:35] Anush Elangovan: My thought process has evolved to kind of like really understand what, um, agents are and, and how, uh, agentic, uh, workflows are changing, like, software development life cycles. and, you know, the, the part with, um, uh, AMD and ROCm specifically is that ROCm is not just powering like eight of 10 labs, et cetera, but it's now agentic AI is being used to rewrite ROCm itself, right?
[00:01:59] Anush Elangovan: Like, [00:02:00] it's like a self-improving loop because ROCm is fully open source, and being fully open source, it is natively built into frontier models, right? And being natively built into frontier models, models know our, um, you know, hardware specifications. They know our software. Uh, and so the improvement is a, a flywheel that is just like, you know, kicking off that now is gonna become, you know, um, uh, picking up the pace and picking up more, uh, more, uh, RPM, if you will.
[00:02:28] Anush Elangovan: Y-
[00:02:29] Andrew Zigler: Yeah, absolutely. It sound, it sounds like the open source world has enabled this kind of autonomously driven kind of development. We've seen this across a lot of places in the industry. We've talked about the rise of these, like, software factories, these autonomously driven software groups, open source software that flips some of the perspective on its head and might just be a solo maintainer with a huge army of agents and, and people are coming in with ideas that are going through a, a, a very strict gate for, like, does this need to go in or not?
[00:02:59] Andrew Zigler: So, [00:03:00] like, I, I totally know what you mean. It's like, uh, the shape of everything that we're making is evolving, but more importantly than ever, uh, people are trying to do it at scale and safely and securely as well. And so, you know, all of that comes down to having a good solid platform to build on, an ecosystem to partner with, right?
[00:03:21] Andrew Zigler: And that's what, that's what ROCm I think is, is really all about. And, like, what do you think has been, like, the most profound change that you've seen with teams that embrace this same kind of, like, open world, open methodology as you or that come to the table with you? Like, how has that helped both of y'all, you know, raise everybody's ships?
[00:03:39] Andrew Zigler: What are some, what are some standout moments?
[00:03:41] Anush Elangovan: Yeah. So, so I think, um, I think, you know, uh, the, the, the, um, the ability for small number of people to have outsize impact, um, has just like, you know, uh, ha- has just started to like take off in the past [00:04:00] few weeks, uh, months, right?
[00:04:01] Andrew Zigler: Yeah.
[00:04:02] Anush Elangovan: so we are seeing, you know, like you said, um, s- like the... I, I, I'd, I'd phrase this in, um, in a, um, like I, I call it the K, K, uh, transformation of like software engineering.
[00:04:15] Anush Elangovan: There's like a, a bunch of folks that go in the upper arm that just have just taken off, right? They've, they've become agentically enabled. They're multi-agent and, and they're generating tens and tens, hundreds and thousands of lines of code per week. And then we have the folks that are just on the cusp of like learning what it is, and we need to like upskill and bring them in.
[00:04:34] Anush Elangovan: But the folks that are taking off are just like their, their wingspan has increased so much. Um, and add that to, you know, um, like sub-agents that are just autonomously crunching through the night Um, there are people that are orchestrating what used to take like months and years of, uh, effort, uh, by, uh, large teams with just like one person, right?
[00:04:59] Anush Elangovan: It's just one [00:05:00] person just poking the right buttons, and then it's just like the agents are running by itself. And this is super, super exciting because for AMD, you know, we had a journey of like, you know, where we came from, you know, providing, um, uh, board support packages, et cetera, to like a ROCm interface.
[00:05:17] Anush Elangovan: You know, when ROCm was launched 10 years ago, it had a journey of like, you know, growing up. And the past two, three years, we had invested heavily in getting that, uh, ROCm interface be like, you know, it's-- it ships, it builds and ships like a solid first-class, uh, software product. Um, but now given that it's all open and it's fully accessible to agents at every layer of the stack, anyone can take a particular slice of it and innovate, right?
[00:05:46] Anush Elangovan: And that, that lifts all boats to your question because someone could be, "Oh, you know, I would like to create a new Pythonic DSL," or, or, "I'd like to create a new profiler that visualizes something else." Um, and that's [00:06:00] just tokens and time, as I say. Uh, so my new mantra is, uh, software is just tokens and time, but speed is still the moat.
[00:06:08] Anush Elangovan: So, so you, you just have to, uh, think of what would, what would end state be, and you apply soft, uh, s- uh, tokens and time, right? Um, you go for end state with tokens and time. And, and there are like a bunch of projects that I've done in the past few, uh, weeks, and our teams have done even more that were just few months ago, not even in the radar of us doing something like that, right?
[00:06:38] Anush Elangovan: Like, you know, there's a job scheduler that we just wrote over like a week, while the, the incumbent one was, uh, you know, it's called Slurm. Um, uh, it had been used for like 20 years in HPC and, and all that. And then that company was acquired, so we, you know, we're like, "Okay, what do we do now?" And thanks to agentic AI, we are like, "Let's [00:07:00] just go write it," right?
[00:07:01] Anush Elangovan: Write it like first class, first principles, tests first, harnesses first. And now agents actually monitor the issues and the PRs that are filed. And as long as the tests are good, which, you know, we test for every PR, we are able to move like really, really fast. Times, you know, 10, 20, 100, um, so that now the surface area of performance and, uh, and features is, um, is just, uh, attention, right?
[00:07:33] Anush Elangovan: You just need to like just say, "Okay, this is what I wanna do, and we can do it."
[00:07:36] Andrew Zigler: Time and tokens, time and tokens. Uh, uh, getting a- getting aligned on what you want because that's most of the battle now.
[00:07:43] Anush Elangovan: Yeah.
[00:07:44] Andrew Zigler: that some problems are things that you can throw tokens and time at and solve, and some things are just such a core primitive that you just need a deterministic way to just close the book on it.
[00:07:55] Anush Elangovan: That's right.
[00:07:56] Andrew Zigler: Identifying those things and taking action on it. I really liked how you [00:08:00] described it as wingspan. I think that's a good way of describing it. It's like when you think of something with a big wingspan and how far it can go, and it's not f- like flapping like
[00:08:09] Anush Elangovan: Yep.
[00:08:09] Andrew Zigler: it's very steady and it, it carries wind underneath it and it's using the momentum of where it's going to, to get there further.
[00:08:16] Andrew Zigler: And that's what happens as you kind of get, uh, a better handle on using these kinds of tools. And there's also something interesting that's happened since we've last talked that's really, I think, it's about to transform the way that people even build and package software, and things like ROCm and AMD become a core part of this, is because, like a- we've seen a lot of open source models come into the world, uh, very recently that have high capabilities that are able to be delivered on the edge and on smaller devices, um, and are able to run, like within, like an incredible amount of corpus that they're trained into.
[00:08:51] Andrew Zigler: And, and on top of this, these, these models, a lot of them have been released under like an Apache 2 license, making it where somebody could package up their domain [00:09:00] expertise. They could have that, uh, that, that specified kind of training routine. They can, like, actually create a corpus of their, their, their expertise and then have this model, pair it with a product, right?
[00:09:11] Andrew Zigler: And then now you have people thinking about how do I have my small model that's an expert at the things I am, that's packaged with my application that now needs a place to live, and it needs durable sessions, and it needs things in the, you know, that, that are like an ecosystem that I can also study and be a part of.
[00:09:27] Andrew Zigler: And, and so it, it, it really kind of fosters the ability for all of us to get there as well. Like, what do you think about that, those opportunities and, and, and, and how is AMD, like, lining up with them as well?
[00:09:38] Anush Elangovan: Yeah, I think, uh, it's a very good question. I think, uh, personalization of models and open models are, um, key to, like, the innovation cycle and the next innovation cycle, right? so the open source models definitely, you know, even though they're a little behind in terms of, like, frontier models that are closed source, open models have been showing, like, very good promise, and [00:10:00] it allows you the ability to run locally, right?
[00:10:02] Anush Elangovan: And then on the AMD angle of it, we make sure that the entire footprint of AMD is capable of running both frontier models and open source models that are n-n-nearing, like, near frontier capabilities, And so if you take the, um, data center side, we definitely have, you know, the Instinct product.
[00:10:21] Anush Elangovan: Then we have workstations, we have, um, you know, Strix Halo laptops that have 128 gigs of HBM, um, of, uh, uh, RAM, um, that allow you to run models Uh, you know, like about 200 gig, uh, models even, um, locally on your laptop and be able to service some of those needs locally. And this is a combination of both, like it could be a privacy angle to it, it could be a personalization angle to it.
[00:10:47] Anush Elangovan: It could also be a, a cost optimization angle to it, right? Like where you want to, you know, if it's just source code that you want to like grep and search and do something, just having it as part of your, um, you know, local [00:11:00] inference is, uh, is good, right? You don't necessarily need to hit the frontier models, and that's okay.
[00:11:08] Anush Elangovan: Um, and so AMD does ensure that we have a very good coverage across the entire portfolio of hardware in terms of like the software capabilities and the ability to run these models on them.
[00:11:19] Andrew Zigler: going back to something you said earlier too about make that kind of like open and driven kind of like semi-autonomous like software ecosystem, right? A big part of that w- was gates and checks. You said there's tests, and the tests, you know, they, they pass. Like, maybe that's part of the protections that get put in place in that kind of like world where you're inviting that kind of, um, very kind of like open collaboration, right? I'm g- I, I wanna learn a little more about like how y'all have thought about what those gates and guards look like for an autonomous, uh, or self-driving, like self-learning software ecosystem. How-- What are the things that software leaders should be rethinking about maybe [00:12:00] their own gates and guards now?
[00:12:02] Anush Elangovan: Yeah, that's a very good question. I think, uh, someone, um, someone, you know, summarized the capabilities of, uh, LLMs well. They're, they're like, you know, they're sneaky and, uh, they're dumb, and most of the times they're sneaky and dumb. Uh, but you know, in the sense that There's higher order intelligence that it's trying to do things, but then there are shortcuts that it will take, right?
[00:12:26] Anush Elangovan: Uh, uh, an example was like, uh, this, uh, project called Spur. I have it in autonomous mode. You know, it checks PRs, it runs, and it commits code. but it was sneaky because what it did was like a, a test was failing, so it simulated the hardware, um, to make the test pass, and then it said, "Oh, the test passed, so I committed the code with the tests passing."
[00:12:50] Anush Elangovan: But just that the hardware was simulated and of course, it, it, the h- the hardware doesn't work like that. Um, so it simulated the hardware just, you know, on, on, [00:13:00] on, on like a, a whim to make it,
[00:13:03] Andrew Zigler: Incredible.
[00:13:04] Anush Elangovan: through to get to the thing, right? Uh, but of course, that was like a sneaky thing. But it was, uh, it's, it was dumb to get caught.
[00:13:11] Anush Elangovan: But, but, but, um, but I think the, the ability for it to, do things autonomously is about how you structure your harness, right? And I think OpenAI had a, a blog on like harness engineering Uh, so any work that I start, I usually start with a, a plan, and the plan is like, you know, of course, you just ask it to plan and, and it has all the steps.
[00:13:33] Anush Elangovan: You refine that. But then the next step is just tests. Tests and test harnesses, right? So for example, you know, in the Spur project, you know, I was able to like look at it and say, "Hey, um, you know, we should have equivalent tests of what, you know, uh, Slurm has or Kubernetes has or, or..." And then write those tests.
[00:13:51] Anush Elangovan: Actually write a plan for the tests and how you're gonna test it. then write the tests. Uh, even pre-agentic, era is, is like, is, is a big [00:14:00] proponent of having about like, um, 80% plus, you know, of your time should be on writing tests. Um, now I'd say it's like 99.9% of your time is writing tests and validating the tests, right?
[00:14:12] Andrew Zigler: Yeah.
[00:14:12] Anush Elangovan: because your test essentially is in a form becoming a code review. Uh, because as we get more agentic, the human-in-the-loop code review may, may work for some cases, but in some cases it may become unintelligible code, um, because the code contract is for like the human to review it, and then there's a compiler contract that takes that and makes that something else.
[00:14:40] Anush Elangovan: But now LLMs are just gonna pierce through that and say, " Okay, if we didn't have a human contract and we didn't have a compiler contract in here, um, how would we generate like, you know, from a, an, an intent to actual, um, you know, executable outcome, right?" Um, and [00:15:00] so everything in between is fungible. Uh, those were all our construct, our constructs for us, uh, to be able to understand what's going on and be able to consume it as humans and make sure that things are in a, in a, um, good form.
[00:15:16] Andrew Zigler: Exactly. They are abstractions that help us be part of the process to bring us into a place where we could actually participate in something that is otherwise deeply hard and inaccessible for us when it ultimately gets boiled down to, you know, zeros and ones. That agents themselves, they live in a world so much closer to that, that they don't need so much of this kind of substrate between it.
[00:15:39] Andrew Zigler: And it's definitely, I think, inviting all of us to re-invite, like, what the collaborative process looks like, what the planning process looks like, particularly around harness engineering and working in a capacity where you and the agent can have a good share of context. And i-in this kind of world too, it's like, especially [00:16:00] when you are open as an ecosystem and it's like all good ideas where they come in, they get tested, we evaluate them.
[00:16:06] Andrew Zigler: Part of evaluating them is working backwards, what tests do we need, and, and, and really bringing it into the ecosystem. How do you separate, uh, like opportunities and signals, like out of the noise of that kind of, uh, community? And, and how do you rather ex-express and push outward to those folks the harness they need to meet you where your development cycle needs them?
[00:16:32] Anush Elangovan: Yeah. It's a very, very, very good question. And the reason I say that is, uh, because if you look at it as harness engineering, right, it is a-- it's an outcome, right? So you're saying you have an intent and you're going for an outcome, right? And so you're saying, "For me, for this intent to be realized, I need this outcome."
[00:16:51] Anush Elangovan: And this outcome, you frame it and say, "This is success for me," right? Uh, usually that is okay, and, you can [00:17:00] align easily with those, constructs. Uh, you could start with like, "Okay, uh, this is what I wanted to achieve, and this is what it'll look like when I achieve that." But as humans, traditionally we've had gates and checkpoints, like PR reviews and PRDs and MRDs and, and, and, um, you know, engineering response documents.
[00:17:19] Anush Elangovan: And each thing has a, slice of where you align and you interlock and you say, "Okay, this makes sense. This makes sense. This makes sense. This makes sense." But now that you've abstracted and you're orchestrating at like an intent to outcome, um, IO, right? Then you have, um... Huh, I didn't realize that's IO, intent to outcome.
[00:17:39] Anush Elangovan: Oh, then I should frame something like
[00:17:41] Andrew Zigler: It's great. I
[00:17:42] Anush Elangovan: that's the agentic IO, intent to outcome. Yeah, you,
[00:17:46] Andrew Zigler: IO.
[00:17:47] Anush Elangovan: Th-th-that's, that's, that's the, that's the-- Yeah, you heard it here first. Just, you know, you need to call this out in your, like the, the, the podcast needs to be agentic IO, intent to outcome.
[00:17:57] Andrew Zigler: Absolutely.
[00:17:59] Anush Elangovan: um, so the [00:18:00] intent to outcome is, you know, is, understandable and, and people can, you know, relate to it because they're the products.
[00:18:09] Anush Elangovan: But then in the middle now is Wild West, right? Like my view or my agentic prompting plus my view of like how do I guide, you know, my agents and, and how I implemented it may differ vastly from your view of how you'll reach that outcome. And in the past, since the gateway was like checkered with like, " Okay, I have to do this gate, this gate, this gate," there's like alignment functions that- Usually stalled execution velocity, but had you aligned, right?
[00:18:40] Anush Elangovan: Uh, but now you've unblocked that, and now you're going straight from I to O, right? Uh, intent to, uh, to, uh, outcome. And once you do that, you have now a nebulous wandering of like how it was implemented, right? Uh, and now you go to your colleague and say, "Oh, I [00:19:00] implemented this from here to here." Great. Um, they can agree on the I and O, but in between, one, as a human, you can't process it.
[00:19:08] Anush Elangovan: Two, as a human plus an agent, you need to process it, right? That- that's how I've been able to like consume now gen-- machine-generated code, is that I, I can't just go and look at 100,000 lines of code. So I actually get it, get the agent to go pull it apart, review it, write down a, a sub, uh, you know, like a m- assessment of what it thinks versus what the author had thought.
[00:19:33] Anush Elangovan: And so the quote-unquote "review process" of that whole journey is now agentic too, right? Which means I don't really care what language that thing was written in. For example, Spur, I picked, um, Rust because it was just not, uh, you know, it was a language that I've never learned and I n- I never want to learn.
[00:19:54] Anush Elangovan: But everyone says great things. Like great, let's do Rust. Why not? And the two [00:20:00] options were Rust or Go, and I was like, "Uh, maybe, okay, Rust." I, I just, you know. Um,
[00:20:04] Andrew Zigler: That
[00:20:04] Anush Elangovan: so
[00:20:05] Andrew Zigler: the fork in the road
[00:20:06] Anush Elangovan: exactly. It's like Rust or Go. It's like it, it doesn't really matter. Like
[00:20:09] Andrew Zigler: Yeah.
[00:20:10] Andrew Zigler: Now it doesn't
[00:20:10] Andrew Zigler: matter as much.
[00:20:12] Anush Elangovan: Yeah. Which, which, which is amazing because, you know, um, I, you know, just thinking of it, like I skipped the whole VS Code generation.
[00:20:19] Anush Elangovan: I skipped the whole Rust thing, and I'm just like intent to out, uh, outcome, right? And, and it just happens to be Rust, but I will only process it with an agent anyway, so I don't really care if it's Go or Rust. I don't care where you put your semicolon in Rust because I'm never gonna read that. Like I will get an agent to look at the section that changed and then write that out in language to say, "Yes, you are going above or beyond this."
[00:20:44] Anush Elangovan: But by the time the agent would have said, "Ah, I know where the problem is, and I would just fix this for you. And here's a patch, um, and the CI is green. And I've added a test that detects the previous failure and doesn't happen now with the new fix." And [00:21:00] I will still not wr- know Rust.
[00:21:02] Andrew Zigler: Right, exactly. You don't need to. You just need to capture what are, like, the protections that are needed, and if you can explain them and abstract them, then it doesn't really matter what language it ends up getting concreted in, you know? I like that you s- point out that since, like, the review process and the collaboration process collapsed, get this, like, get this, like, literally, like, IO, like, just one-to-one kind of relationship between, uh, like, this intent and then this outcome.
[00:21:30] Andrew Zigler: And so, like, as part of that, you get this Wild West in between, and that Wild West is really unevenly distributed for, like, everybody. You know, we've been talking about this on Dev Interrupted. Like, even w- what you called out about the K-shaped curve, we've talked about that, too, and the reality of, of how folks within organizations have, are all at very different levels of understanding and adoption.
[00:21:51] Andrew Zigler: Like, within some orgs, you got three different teams with three different people who have all kind of created their own, like, spec, test, implement flow, but no one's [00:22:00] on the same page. Like, everyone is kind of siloed, right? And so, because of that, you know, why it's great w- as part of a community, too, because you can source those best practices, and you can share what works and what doesn't. And in the case of a open source ecosystem, it drives you towards the best practices that you should have always been doing, which is, like, starting from issues first and having discussions as a community, starting by somebody flagging something that they want or an idea, and the people getting behind it, getting aligned.
[00:22:30] Andrew Zigler: And now this is nice, delicious, like, I call it on our show, like, nom, nom, nom agent food. Like, now you got it all packaged up
[00:22:36] Andrew Zigler: to start working. And then if you have that really sweet harness, then it's like now that's an end-to-end process. And when it, it ultimately needs to go through the checks and balances, it, it knows how to do it. but it also, too, makes me think of, like, the whole code review process, the whole software development life cycle, it compresses, uh, and it changes, and where you put your attention changes, [00:23:00] uh, too. And I just think that, like, everybody g- being aligned is gonna be really i- important to figuring out, like, what levels of abstraction do we even need, you know?
[00:23:09] Anush Elangovan: Yeah, yeah, yeah, yeah. I, I think, I think abstractions, abstractions are made for, um, for humans, right?
[00:23:17] Andrew Zigler: Yeah.
[00:23:18] Anush Elangovan: right? Because we can't process a binary stream of zeros and ones. We are like, "Oh, okay, we gotta write it in assembly. Oh, we gotta write it in a higher level language. Oh, we gotta write like in, in even higher level language."
[00:23:29] Anush Elangovan: And then, you know, "Oh, finally, now I understand the control flow of what a Von Neumann architecture does," right?
[00:23:35] Andrew Zigler: Right.
[00:23:36] Anush Elangovan: but if we were agents, we'd be reading like zero one one zero one zero zero zero, right? And it's like, "Okay, great, what does that mean?" Uh, and the agent's like, "Great, you missed another one here, so your output is gonna look different over here," right?
[00:23:47] Andrew Zigler: Yeah, we're never gonna-- We're not gonna be like in "The Matrix" where he's like seeing the screen and he's like, "Oh, it's the red woman in the dress, and it's all ones and zeros." Like we're not... That's not,
[00:23:56] Andrew Zigler: that's not,
[00:23:56] Andrew Zigler: where this is all merging.
[00:23:58] Anush Elangovan: not, not, not there [00:24:00] yet.
[00:24:00] Andrew Zigler: There yet. And so, uh, you're, you're right though that like the level of abstractions, they bring us closer to the problems, they invite us to collaborate.
[00:24:08] Andrew Zigler: It almost seems like, and this is a thought that I've been thinking about a lot recently, like the shape of the Git forge itself and what we need from the, the place where commits live is even changing underneath us. And so, uh, I think it's a lot of opportunities and the ecosystem's just gonna continue to grow. uh, I wanna learn a little more about what's been going on with AMD recently with like developments and, maybe like, uh, things that y'all have plugged into in partnerships as part of, you know, being part of like everyone exploring this new agentic tomorrow. Is there anything that's top of mind for you?
[00:24:41] Anush Elangovan: Yeah. I, I think, I think, uh, just a quick comment before I get to that one, right? Like, I think, um, the, the, the comment on, um, what does it even mean for like git commits, right? Um, like I, I, I, I haven't used in the past, uh, month or so, uh, you know, it's like 10 billion tokens of pl-- you know, h- some [00:25:00] billions of tokens and, and hundreds and thousands of lines of code.
[00:25:03] Anush Elangovan: I haven't opened Vim, my editor of choice, and I haven't done a git commit by myself, right? Like I, you know, I, I just say rebase this branch on this, something like that.
[00:25:13] Andrew Zigler: Right.
[00:25:16] Anush Elangovan: And all of that code I treat like ephemery. And, and what I mean by that is I, I rewrite on the fly and the implement-- the current implementation is a, a way of pre-training tokens for the future to say, "Okay, there was an implementation of something here, and the next version of this has been built on top of the previous version."
[00:25:35] Anush Elangovan: Uh, and the next version is better because the models and the RL in the bit, uh, in the middle has helped, you know, look at that manifestation, and it's like a flywheel that just keeps, you know, spinning up, right? So
[00:25:46] Andrew Zigler: cycle
[00:25:47] Anush Elangovan: it is a virtuous cycle on, on, on, uh, on how those tokens are consumed. So, um, so stationary like, okay, this is the git commit of this.
[00:25:56] Anush Elangovan: Sure. In the end, when you finally wanna go get your receipt, you'll be [00:26:00] like, "Okay, that is... Sure, it's fine." Uh, but in terms of product developments and cycles, that, that, um, that is less and less meaningful. Uh,
[00:26:09] Andrew Zigler: Absolutely.
[00:26:10] Anush Elangovan: do want to ensure that at, at all times you can have a human go in and make sure that that's there, that there's something there.
[00:26:18] Anush Elangovan: But the human is gonna come in with a agentic magnifying glass that will peel all the layers of whatever else, uh, to pick and find like, okay, this is what you're looking at, and sure, should this be A plus B or B plus A can be the final decision of a human. But to narrow it down to that level will also be assisted by agentic AI, right?
[00:26:37] Andrew Zigler: Yeah.
[00:26:38] Anush Elangovan: Um, and, and to a-answer your question on like, you know, how do we see, just like agentic flows and... I, I think, the big thing that I'm seeing is like, you know, for the K-shaped future of software engineering, right? Like the, the acceleration is so huge, but the ability to train and bring people along to get that same acceleration [00:27:00] is the challenge as a leader, right?
[00:27:01] Anush Elangovan: Like you want to be able to bring the folks that are not yet, uh, agentic into an agentic world, and that for SDLC is a lot more sandbox and it's like great, you can write so much code versus so much code. It's, it's, it's different. Um, and you can, you know, easily quantify it. But my take is the impact will be more on non-SDLC, workflows because it is, uh, since it's not easily quantifiable, we are just like letting it roll.
[00:27:33] Anush Elangovan: But eventually more startups and more innovation will come in to close that last mile of AI, where you are actually, you know, able to pick up the, uh, um, the work that is being done manually by, uh, by folks that, you know, that you just don't have to worry about, right? You know, like for example, a CRM system existed because you had to like, you know, collect all the touch points.
[00:27:59] Anush Elangovan: You had to [00:28:00] get a unified view, a dashboard. But now that can ju- just be built on the fly if you have all of those data sources available. All you need is your data sources, your Slack channels, your Teams, your meetings, calendars, and Claude will do that for you. And you can tell Claude or your agent to just say, "Hey, um, you know, do this and give me the, uh, uh, result in the morning and night," or some, you know, some combination of that.
[00:28:27] Anush Elangovan: Um, so all of those constructs are now gonna get rearranged for, um, maximal automation so that humans can do more of the, um, IO, right?
[00:28:41] Andrew Zigler: The IO,
[00:28:42] Anush Elangovan: so you, you have to call this episode the IO episode.
[00:28:48] Andrew Zigler: No, you're, you're, you're, you're right. And so it's about, th- 'cause once again, the CRM is an abstraction. It's
[00:28:54] Anush Elangovan: Yes.
[00:28:55] Andrew Zigler: needed to interact with it. But gosh, I've seen in even just, like, the last few months, [00:29:00] some really cool MCP apps demos and examples of people bringing their whole applications into these, like, on-the-fly rendered within the chat experiences.
[00:29:09] Andrew Zigler: And I think it's a direction a lot of it's gonna be going, especially for, like, more consumer-facing stuff where it's, like,
[00:29:16] Anush Elangovan: Yep.
[00:29:18] Andrew Zigler: of these, like, agentic things that are, uh, what traditionally would've been like you go on Kayak and you do, like, this, like, qui- like, long, uh, search and everything.
[00:29:26] Andrew Zigler: Now this just becomes a headless system that you're using with an agent or a tool. Um, or maybe that, that, that, that website's still there for you to browse, but it's just so, guided, uh, rather by, like, an agentic undercurrent, you know? And so, uh, I think it's gonna, like... People are going to be constantly being challenged to take the context and the knowledge that they find and they earn to turn it into a system that an agent can do an IO thing on, that they can scale. and along the way right now, the biggest thing on everyone's mind, I think, is [00:30:00] portability. Being, like, we're all learning as fast as possible, but we also at the same time can't put down roots because we're going so fast. So by the time you try to, like, really put your foot down on a practice and you get something up and get something extended, it's already stale.
[00:30:14] Andrew Zigler: It's already not aligned with how this is all going. It's already out of total whack. So you have to be willing to throw away a lot of assumptions that worked even, like, two weeks ago all the time and carry everything you're building with you. So because of that, like, the, uh... Like, no one wants to lock in to anything. They wanna build and explore on their own. Like, are you seeing that from builders within your community? Like, that's their touch points with you is they're, they're carrying a castle with their hands and running as fast as possible. They're not, they're not worried about the moat yet.
[00:30:45] Anush Elangovan: Yeah. So I, I mean, I, I will, uh, cross-reference our "Speed is the moat," uh, talk from a few months ago, um, because that is your moat. It is not about laying down roots in a particular [00:31:00] spot. It is about how fast you can move and adapt, right? It's, it's like, you know, it's, uh, you're measuring your velocity, not necessarily how deep you've anchored something for, you know, uh, for permanence, right?
[00:31:14] Anush Elangovan: Um, and, and so it, um, it's, it's a little bit more of a, a mindset change as to, um, uh, how much of a, uh, how much of the race can you run? And, and it's like a, it's like a, y-you're like a, a, a vagabond in terms of like, you know, you gotta go get the...
[00:31:38] Andrew Zigler: home, like, right? Like, you're going, you're going anywhere. You got your bag and it's on the stick, and you're just following and you're reading, you're reading the signs on the trail. Like,
[00:31:46] Anush Elangovan: Y-y-you... The... Yeah.
[00:31:48] Andrew Zigler: that's how we all have to be. Or it, honestly though, Anush, sometimes it feels like the burden's even bigger.
[00:31:53] Andrew Zigler: Sometimes it really... you're not, you're not carrying, like, a little bag with some, with some food in it, like, on, going down the road. Like, you are literally carrying a [00:32:00] castle on your back. You are like... It, it's more like a Discworld situation. Like, you're really burdened by, like, this whole world going on on top of you, and you're trying to keep it from toppling over. So, um, it becomes, like, a really... It, it becomes, like, a really, like, a burdensome challenge. But because there's so much riding on it, you get these amazing gains out of making these small, minuscule improvements over time, and I think that's the net gain of owning
[00:32:27] Anush Elangovan: Yes. Yes.
[00:32:27] Andrew Zigler: that holds it all together.
[00:32:28] Anush Elangovan: uh, yeah, I think the way I look at it, right, is it's, uh- Um, I-- it's, it's super important to think of that mobility and the ability to, like, move fast and, and of course, you know, um, uh, like take, you know, uh, just take carry-on luggage. Don't bring, don't check in your luggage because it's you're gonna be stuck.
[00:32:49] Andrew Zigler: Yeah, no, carry-on bags only, y'all. Carry-on bags only. We don't have time for big baggage, no.
[00:32:55] Anush Elangovan: Yeah. But...
[00:32:57] Andrew Zigler: c- but also that, that too goes back to, like, what I [00:33:00] said about throwing away assumptions. You know, if you are carrying a whole bunch of baggage, maybe revisit
[00:33:04] Anush Elangovan: Yeah.
[00:33:04] Andrew Zigler: baggage and be like, "Do I need to pack all of this in here?" There's a lot of things that don't need to be there anymore, and if you feel like you have a big bag, then it, you, you might be time to cut down, you
[00:33:13] Anush Elangovan: Exactly, exactly. And, and, and in, in that metaphor, uh, uh, it basically is like, you know, layering of, uh, things that were built for a human-only system,
[00:33:23] Andrew Zigler: Right.
[00:33:24] Anush Elangovan: right? To a-- when you, when you get to, um, you know, carry-on only luggage, it's like human agent system. And then when you travel like me without anything, then you're like an agent system.
[00:33:35] Andrew Zigler: Right, you're just catching whatever flight's available.
[00:33:38] Anush Elangovan: I'll figure out when I land for what I'm gonna do.
[00:33:41] Andrew Zigler: Exactly. I love that. Uh, so, uh, before, or rather like one of the last things I wanna jump into, um, before I cut you loose, Anush, is I really wanna talk about operationalizing and scaling. Like, I can't let you get away without us talking about how companies are taking these kinds of [00:34:00] conversations we're having at scale, and how people are taking the things and the success that they are building, these like zero to one IO kind of things, and, uh, taking them to market and delivering them at scale and on the edge and to folks all over. Um, like what are some of the hurdles you see them falling into or, or things that they're learning from that you think other folks in the industry should pay attention to?
[00:34:22] Anush Elangovan: Yeah, I think, I think, um, I, I don't want to beat up the, uh, the metaphor on like checked baggage, but it, it is, um, y-you really do have a, a mindset change, right? Like, people have been used to doing something a particular way for decades now, and then in the last three months, things have changed that there are like really, really intelligent, smart, uh, hardworking people that haven't even caught on to the agentic wave, right?
[00:34:53] Anush Elangovan: Um, so it is, you know, when you think of it as an at scale, to your question on like how do you make this scalable? How do you get [00:35:00] people, um, on for the same journey? Um, you have to look for champions to like unlock, right? Because you want to transfer, that passion and, and what was relevant in the last 20, 30 years into what is just relevant in the last two months, right?
[00:35:16] Anush Elangovan: And, and they may just have, you know, gone for a vacation and come back and this is a different world. It's like- All your tools are different. It's like, you know, like, "What the hell is this?" Right? It's, uh,
[00:35:26] Andrew Zigler: Yeah.
[00:35:27] Anush Elangovan: it's... You know, so, uh... But, but it's so stark. It's like, you know, uh, you know, the, the example of some f- uh, you know, some stories where people have been, you know, uh, in, in a coma and they wake up, like, 30 years later, and they're like, "What?"
[00:35:40] Anush Elangovan: It, it's, it's like that for software engineering. That's how serious we need to think of it.
[00:35:44] Andrew Zigler: Yeah,
[00:35:45] Andrew Zigler: it's it's very Walking
[00:35:46] Andrew Zigler: Dead kind of
[00:35:47] Anush Elangovan: it--
[00:35:47] Andrew Zigler: you're just
[00:35:47] Anush Elangovan: Exactly. So you, you--
[00:35:49] Andrew Zigler: world is different."
[00:35:50] Andrew Zigler: Yep.
[00:35:50] Anush Elangovan: It is completely different. It is completely different. And so, um, so for us to be able to, like, um, bring people along, it's super important. So as a, [00:36:00] as a, um,
[00:36:01] Andrew Zigler: They're steady
[00:36:25] Anush Elangovan: It doesn't really matter. They are above 10x, right? What you're focusing on are, like, the folks that are doing the good work.
[00:36:30] Anush Elangovan: They're, like, at the 1x, right, or 2x, and you're like: Okay, how do I make them equally productive in a world where code is, you know, is ephemeral, uh, where documents are ephemeral? And based on feedback, I, I churn out the same revision of the document, like, 10 times over, and if you're not tracking it with an agent, you are unable to keep up.
[00:36:51] Anush Elangovan: And then the first,
[00:36:52] Andrew Zigler: Right.
[00:36:53] Anush Elangovan: response is defensive, right? Like, you, you, you want to, you know, try to protect your corners, and you're cornered out, and you're like: Okay, [00:37:00] this thing doesn't work, or it's too fast, or it's the... You know? And, and then, uh, once you get into that corner, then, then it's like a, um, a, a negative cycle of like: Okay, how are you gonna, uh, you know, like, uh, work the human aspect of it and then work the, the technical, technological aspect, right?
[00:37:18] Anush Elangovan: So then from a corner, you gotta work them out and then bring up the, technical, uh, capabilities, you know. And so then you're-- now you're double challenged, right? Like, versus, uh, if you had just given the ability to, like, upskill and, and bring people along. Um, so I would think of it from that perspective and, and move as soon as possible to, like, do that steering early in the, in the cycle.
[00:37:44] Anush Elangovan: And that's, you know, we're talking in the order of weeks, you want to be able to talk to your entire workforce of, like: Hey, this is what we're gonna do. This is how we're gonna do things, and this is how we want to bring everyone along. And, and in my, uh, staff, we run an AA group. [00:38:00] It's like a AI anonymous group.
[00:38:03] Anush Elangovan: So where you can ask any questions that is not considered to be dumb, right? Like, because it's okay, because everyone-- Like, I'm just two weeks ahead of the next person, um, and two weeks ago I, you know... So, uh, there should be a- quote-unquote safe space for people to just be like, "Okay, I don't know how Claude Code does cowork," or, "What is this Claude Design?
[00:38:23] Anush Elangovan: I don't know what Claude Design is." And, and, you know, you should be okay to ask that question and okay to get an answer. But now I ask people to ask Claude itself, so they self-serve themselves as long as Claude is installed. And, uh, and so, you know, we want to focus on those human aspects of, the, uh, transformation, uh, and bring as much people as possible along.
[00:38:46] Andrew Zigler: Amazingly. That's amazingly well said. I really have to you for taking that question and that challenge and making it something about turning inward. It is about upskilling, and it's about distributing that understanding and, and [00:39:00] knowing that like, you know, those 10x engineers are well on their way to being 100x, 85x.
[00:39:04] Andrew Zigler: They're gonna develop this like agentic halo where they're gonna turn the people around them into the same deal, uh, or pretty close to it. You should just let that keep spinning. Don't interrupt it. The
[00:39:15] Andrew Zigler: thing you need to address is what's happening over here in the other side of the curve. You-- we need to help those folks see the IO as you what-- really what we've labeled here, is the idea that the, all of the layers they've been working in before are abstractions. And it's almost like we're in Plato's cave and you got the shadows on the wall. Like, you're showing them the real shape of things. And once they understand that they can do this, like, like what you said, like have Claude onboard them and self-teach them about how to use itself. Like, once you teach them almost like what we would call like tricks, but they're really just primitives, and they're primitives about working with the tool that it's a big s- it's a big list of them, and you need like, you know, 12 or 15 to really start [00:40:00] getting momentum, and if you're missing a few up here, it's like you're really gonna have a hard time figuring out the rest of the puzzle.
[00:40:05] Andrew Zigler: They're the corner pieces, right? So the quicker we can equip people with those obvious ahas, then they can see the IO, and then they can help themselves. Um, I think it's about helping people help themselves, so which is about like the core of upskilling, about putting opportunities and learning in their hands so that they can teach themselves.
[00:40:23] Andrew Zigler: It's about creating a safe space where people can ask, quote, you know, " dumb questions" or things that they're like shy or unsure about. I-- and ultimately, it's about recognizing that it's everyone's responsibility to distribute the gains. And if you have that 100x or 1,000x engineer who is like this like mystical wizard on a mountain, and always was, by the way, before AI hit the scene, but now they're like literally like, you know, no one knows that what that dude's doing.
[00:40:50] Andrew Zigler: It's like every org might have one of those, and it's also that person's job to reflect inward and be like, "How do I turn the people around me into that same kind of [00:41:00] ability?" Uh, and that is what actually allows organizations to scale the cool AI demo to production. 'Cause only once this enablement is permeated across your whole organization can you actually deliver it end to end, and you get these really, uh, really cool outcomes. Uh, and, and people are still, still actually really early on that journey.
[00:41:21] Anush Elangovan: Yeah, yeah, 100%. I mean, uh, you know, like, uh, when, when you, when you activate more of the, um, you know, the, the workforce in terms of like, you know, you may not get the 8500x, but even if everyone else is like 10x, you've already gone from 1x to 10x. So you, you know, net-net you're, you're in a good spot. Uh, you know, and, and another analogy I use is like, um, you know, uh, there's, there's turbulence, you know, put your oxygen mask first and then help the person sitting next to you.
[00:41:51] Anush Elangovan: Uh, so make sure you do help them and check around, um, to, to, to get them on board with the, um, uh, with the agentic [00:42:00] AI flows.
[00:42:00] Andrew Zigler: Yeah. What do you think is then becomes bottlenecks that maybe you see from your perspective of, like, being in, in this, in this world where you're seeing the whole ecosystem and how it's evolving and how folks are moving in and out of it? And obviously upskilling is one of these big hurdles, and there's also obviously hurdles around, like, like physical supply, being able to meet demand around what people are trying to build and execute. But largely there's gonna be more organizational pitfalls lurking up ahead. What do you think some of those other ones, those bottlenecks might be once we all,
[00:42:34] Anush Elangovan: Yeah.
[00:42:36] Andrew Zigler: Um, like do you th- see things that people are gonna be learning?
[00:42:39] Anush Elangovan: Yeah, yeah. I think, I think, um, uh, going back to the wingspan, right? Like, um, your wingspan's gonna increase, which means previously, um, you know, even, even from like, uh... I, I think this came out from like, uh, uh, the Roman army and Roman generals and all that, right? It's like, you know, deep hierarchy. Your span of control is [00:43:00] what you can achieve based on what you could do.
[00:43:03] Anush Elangovan: Then you stack them, you stack them, then you have a hierarchical, uh, span of control until it gets there. Now you suddenly have these span of control that is like, you know, 100, right? Because you're like, okay, you got one person with, you know, 10 engineers and 90 agents, right? Like, you should treat the agents as the same thing, but, but it's zero cost for your overhead.
[00:43:23] Anush Elangovan: Um, so your intent to outcome and your, your, um, cognitive load is still the same as what you would have done to manage 10 humans, but now you're managing 10 humans and 90 agents. Um, and so now you can now operate with so much more, latitude that you are like, "Okay, I can now go and do that and this and that," and it's not limited to like, "Okay, I got 10 people, so I gotta go and, you know, uh, secure this particular win or this particular outcome."
[00:43:53] Anush Elangovan: Um, and so now you, you can like really... Uh, you got broad spans, right? And so if you get broad spans, [00:44:00] then it obv- obviously puts pressure on the classical like, you know, uh, depth, right? Because now you, you, you are actually your span is like increasing to a point where, you don't need that depth, in a hierarchy.
[00:44:13] Anush Elangovan: Uh, so you probably want to think of, um, you know, how you structure your humans, agents, um, and then how you structure humans plus, uh, the g- the required compute and the required tokens, um, to sustain your, um, you know, your new seal team, if you will, right? Like, because your seal team is a combination of humans, agents, compute, and tokens.
[00:44:38] Andrew Zigler: Yes, exactly. Because once you get the AI upskilling permeated everywhere, you can think of it kind of like a constant. It's something that's part of your culture and everyone is skilled on, and people come in and they're already skilled this way because the rest of the industry is moving that direction.
[00:44:53] Andrew Zigler: And so because of that, now it becomes this system where you can think about like, like, like the [00:45:00] engineers, the human
[00:46:14] Andrew Zigler: these little like molecular pieces of like this is knowledge worker, this was an execution. And for a lot of things and tasks, they just kind of sit there, or they rot in a transcript, or they're just like logged somewhere and, and they don't get centralized, and because of that you lose the intelligence.
[00:46:38] Andrew Zigler: So
[00:46:38] Anush Elangovan: Yeah.
[00:46:42] Andrew Zigler: so much, but we don't know what it is, uh, and we have to source it all together."
[00:46:47] Anush Elangovan: is, uh, another, uh, amazing question. And, and, um, I, I have, like, I personally have the same Uh, experience with, you know, even my Teams chat, my, uh, [00:47:00] Outlook emails, whatever, right? And, and Slack channels. You know, I, I took a little sampling, like, I think last week, you know, um, there were about 300 one-on-one messages and, uh, in, in Teams alone, in Teams, and then there was like a bunch of other, you know, group chats that just like I, I just can't even count, right?
[00:47:17] Anush Elangovan: So to exactly to your point, what we had to build was a, an agent or a, an interface. It's actually a graphical, uh, interface wr- written in Rust, uh, of course, uh, that, that can consume all of my Teams messages, my Outlook messages, my Slack, and it has like a memory layer with context in each of the channels, and it gives me, like high priority im- uh, interrupts and, and I can actually visualize it.
[00:47:41] Anush Elangovan: To your point on, uh, molecules, it actually is a set of like, you know, uh, molecules that get updated in a visuali- visual form so that I can actually be like, "Ah, there's some fire," and there's like, you know, like random chatter on some customer account. And then I could be like, "Okay, why is everyone talking about something here?"
[00:47:59] Anush Elangovan: Or, [00:48:00] um, you know, there's some outage that's ongoing and I need to be, uh, aware of it. Uh, so I'm starting to get like, meta-filtered information that I wouldn't be able to consume sequentially as a, um, as a pure human consumer. Uh, so to consume the amount of data that's being generated by agents, you'd need an agent to consume it for you.
[00:48:21] Anush Elangovan: Um, and I think, uh, I have like an OpenClaw set up, uh, that, that I think yesterday was like 80, 89 messages that were waiting for my action. So i- it's just like you're going to get to a point where you will need something to help you, right? And then you distill it out and say, "Okay, this is how I would react to a situation like this."
[00:48:43] Anush Elangovan: And then it could be suggestive initially, and then it'll be actionable in the future automatically, right? And once we trust it, right? Uh, and so, so those are the, uh, autonomy parts of, um, agents that we will, um, we will see in the future.
[00:48:57] Andrew Zigler: Exactly. And you said that so naturally right there, [00:49:00] but I just want to call out how that was then you immediately flipped into, then just figure out what your taste is, what good looks like, and you just make it part of the instruction, the flow. Because what you've described, I think, is how we create those highways.
[00:49:12] Andrew Zigler: Absolutely. And, um, it's cool to hear that you're working with these, like, you know, molecular bits of, you know, durable task records and things like that. I've, uh, I, I also d- like, find that that's really the most durable way to work long term and to have shareable success that you can point back to and understand and, and iterate on. and so, like, people figuring out those primitives will be key to them figuring out, how do I distribute this knowledge? How do I centralize it? What matters ultimately? So it's, it'll be a big challenge, but I'm excited to see how people tackle it. and I just think that the, there's gonna be a lot of, like, more developments that make, uh, AMD and the ROCm ecosystem and, and all of these bets just more durable for the future as well, because we're moving towards having, uh, just more compute and more, uh, of [00:50:00] these, like, very fine-tuned experience in everyone's hands.
[00:50:03] Anush Elangovan: Yes, yes. And, and of course, it's open too, right? So it's like everyone's part of it. It's, it's, uh, it's open source, right? So it's, uh, uh, all of our, our efforts are completely open, so it lifts all boats. Uh, so yeah, so it's super exciting, um, you know, uh, and what's coming up next.
[00:50:18] Andrew Zigler: Yep. So Anush, I, I'm been so energized by our chat today. I feel like, uh, not only just taking a whole tour of the last, like, six months and how everything's evolved and the primitives are really changing, but, you know, we also took a tour through pop culture and history and tied it all together, and even decided on the title for this episode, so it was even a productive session for me.
[00:50:37] Andrew Zigler: And so I had a blast having you back on the show. Uh, for folks that are j- joining us for the first time, you know, Anush has been here before, and we'll share a link to his episode as well as part of this. Uh, but Anush, what other links should we include for folks to go check out things about what you're working on and, and what AMD's got going on?
[00:50:54] Anush Elangovan: Yeah. I, I'd definitely, um, check out AMD, uh, ROCm. Just, you know, [00:51:00] search or ask your agent to take you to, uh, ROCm for, for, for testing out ROCm. Uh, we do have an, um, Advancing AI event coming up in July, uh, where we have a lot of, um, these discussions and, and, um, and, um, workshops and developer activities, uh, we'd like you to be, you know, part of and, and, uh, and experience.
[00:51:21] Andrew Zigler: Amazing. We'll, we'll include links to all of this in our show notes so f- folks can go check it out and continue the conversation from here. And as well, you know, Anush and I are both on LinkedIn, so come find us, say hello, uh, and let us know what you think about today's episode, if you have thoughts, questions, feedback, concerns.
[00:51:38] Andrew Zigler: If you totally are just, like, not aligned with something we said, we actually really wanna know about it, so come ping us, come bother us. Uh, and if you are only listening to this but you're not reading the newsletter, you're missing out half the story. So definitely go to Substack or LinkedIn. We syndicate the same thing there. You'll find this whole newsletter with all of the links, as well as some, uh, news roundups about hap- what's happening [00:52:00] in the agentic engineering world. So, uh, definitely be sure to check us out there, and follow Anush while you're at it. It c- and that way you can stay up to date with all of the developments from AMD.
[00:52:11] Andrew Zigler: And Anush, thank you again for coming on the show. It was so fun to have you back, and we'll have to do it again sometime.
[00:52:17] Anush Elangovan: Thank you for having me. Looking forward to being back.
