Agentic AI: Dissecting the Future of AI Workflows | Memra's Founder Amir Behbehani

0:00The work of an AI engineer is really to kind of build those systems that effectively build systems. I guess if you're building agentic frameworks and then deploying those agents to effectively write code, and then the code is for the purposes of building an application, then you're building the systems that effectively build the constituent systems that build the application. And in that regard, it's sort of more like industrial engineering than it is necessarily software engineering. 13 % of all poll requests are bot created today, and they're creating a unique impact on your SDLC. LinearBee's upcoming research exposes the effects bots are having on your team's developer experience and productivity, and engineering orgs who created a system for managing bot-generated PRs are able to reduce their entire review load by over 6%, while also making drastic improvements in their security and compliance posture.

0:51You want to learn how your team can manage bot-generated PRs and get early access to LinearBee's data report? Head to the show notes to register for our upcoming workshop on September 24th or 25th. Welcome back to Devon Ruptid, everyone. I'm your host, Connor Bronson. And today I'm joined by Amir Babani, a mathematician and AI ML expert. Amir is both a founder and the chief AI engineer at Memra, and he's had previous exits to both Google and Meta. Amir, thank you so much for joining me today. Thank you for having me. Yeah, I'm really excited for this conversation because we've talked a lot about how AI is impacting software engineering from the perspective of engineering leaders, but we haven't really had the opportunity to dive in depth on what's happening as far as the forefront of it and kind of what we see moving forward with someone who's a deep expert in the research side.

1:38So in today's conversation, we're going to have that chance to zero into the future of AI and software engineering and hear from you about kind of the cutting edge of research, plus talk about agentic AI and how that may transform software workflows or other workflows in the world. But before we jump in, I do want to remind our listeners, I know we say it every time, but it really does matter. If you enjoy this episode, please just take one moment to rate and review the Dev Interrupted podcast on your app of choice, Stitcher, Google Podcasts, Spotify, whichever it is. It helps us bring more insightful conversations with leaders like Amir.

2:10And if you really love it, tweet it or Zed it, post on LinkedIn. We love to hear from you. But with no further ado, Amir, let's dive in. Agentic AI is rapidly redefining how we interact with technology right now. Can you start by explaining for our audience what Agentic AI means, just make sure we're all on the same page, and how it differs from prior concepts folks may have heard of, such as RPA or robotic process automation? Sure. I think of agentic AI as blocks of code that have reasoning capability. So they actually have access to an LLM. That's the first thing. Then they have access to long-term memory, and you can think of that as akin to the RAG layer.

2:50So that would be the vector databases plus the graph databases plus other long-term forms of memory. Then they have access to sort of short-term memory that increases their adaptability. And then they have access to tasks on which they're trained. And then finally, they have access to some sort of integration layer. They integrate into enterprise applications or other workflows. And in so doing, unlike, let's say, RPA, because I often get the question, how does this technology contrast with? RPA, these agents are interacting with knowledge as context, as opposed to sort of data inputs. And they can sort of reason their way to a conclusion when they're interacting in so far as these workflows are concerned.

3:38So for example, you can say, write these documents to a database that begs a whole suite of questions. What fields in the database? How did the fields in the database map to this particular document? What do you want me to extract from this document? Write to the database. What database? What's the path database, et cetera? So the agents have the ability on the fly to reason through that imperative and get to an outcome that was so desired by the user who's sort of commanding them to do that particular job. And this is really interesting because I think most of us are used to deterministic code, essentially, where, hey, we expect the output to remain the same.

4:24And this more non-deterministic model of an agent or typically non-deterministic, you get so much more capability as far as reasoning, but the array of results you may receive is more varied. And this becomes particularly interesting for folks in software engineering, because as we've seen at Linear B, we've just released a 2024 bot automation research study that found that more than 13 % of pull requests today are already bot created. And we expect that number to increase with this continued innovation in AI, more happening on the agentic AI side of things. How rapidly do you expect to see fully AI, non-deterministic models come through and actually commit code to a code base?

5:08I think it's happening as we speak. I've tried a myriad of these tools myself and these other frameworks. I'm both a consumer of these things and a producer, and I enjoy using them. They're actually quite fun. I think in a lot of cases, there's a lot of work that is needed for these things to fully automate the development and the deployment of large-scale applications. However, the pace at which that work and sort of that innovation is taking place, I don't see it too far in the distant future when at least small scale, fully functioning applications are developed and deployed through a single prompt with a bunch of the reasoning and sort of the Socratisms that happen between the single prompt and the final outcome taking place in between the command and the outcome.

6:01So I do see that taking place. And actually, that speaks to, well, how do you maintain context? How do you actually break down a particular complex function into constituent sets of tasks, as you would maybe map reduce for a workflow? And for example, what is the right network topology to maintain context and to get a job done? So for example, there are different types of network topologies. A bucket brigade is a very simple network topology. There's a house burning here, there's a lake here, and you just move the buckets in a linear fashion. If you're disseminating orders in sort of a complex organization, you want that topology to be hierarchical.

6:44if access to information is necessary sort of in near real time you kind of want to have a flat or heterarchical topology so identifying a given problem and mapping it to the right network topology so that you can break the tasks down into sort of a constituent set of subtasks and then have agents that are trained on specific subtasks allocated to those subtasks and then have let's say a master agent as a reducer summing the sums. Okay. And so if you kind of pull that together and that's sort of like what's missing in a lot of these application or agentic frameworks that I've tried thus far as a consumer, if that can be pulled off, yes, I see that in a not so distant future, you can sort of command the development and deployment of at least small scale applications.

7:35So my kind of mental framework for this is that we're seeing this shift from something that I think a lot of us have done for a long time, which is, hey, we're going to automate away repetitive tasks to not only are we going to automate away repetitive tasks, but we are going to essentially create agentic employees who are working on your behalf and where you have to really think about the inputs or even those employees. And maybe there's less of the people management, but the information management becomes even more important. Is that an accurate framework to apply here? I think if you look at agents as they stand right now, without mentioning any specific agentic frameworks, but if you just, there's some well-known ones out there.

8:17What they do is they reference a function as a tool. And then that tool, along with several other parameters, is allocated to a particular block of code. So the block of code, as I said, has an LLM. it can have even an LLM router it will have some sort of memory and it will have access to this tool and that tool is making a it's effectively making a function call or it's having access to a function now if you think of these functions as methods that belong to a class and if you can sort of further abstract that the classes sort of fit into larger workflows then you can and this is actually what we're doing so kind of speaking from from experience here we're registering into memory of a master agent, not a single function to which it has access through a tool, but whole classes and methods that map to a particular workflow.

9:17So effectively, you've taken a master agent and you've trained it on a set of workflows. Well, if you think about that in the context of human workers, a human has a role. That role maps to a job. That job maps to a set of workflows. Those workflows map to sets of constituent tasks. And right now, I think what we're saying is that if we can take a workflow and break it down into constituent tasks and allocate agents to those tasks and complete those workflows, then effectively, those jobs are being done by the agents, which then means these agents have a role to play within an organization. And in that case, I would say that these agents are effectively digital employees.

10:04And those digital employees paired with a human employee, say an individual contributor, suddenly that individual contributor has multiplicative returns to scale, as does hiring a manager. If you hire a manager - There's your 10x engineer. Perhaps, yes, indeed. Perhaps, yeah. Yeah, even, and actually think about it from this perspective, if you have the employee right now using, let's say, chat GPT, there's marginal improvement on their productivity. But if the human is taken out of the loop, because that human is maintaining context through the dialogue with the GPT, sort of that's they're sitting around maintaining context, engaging, winnowing down on a set of conversations and therein the outcomes of those conversations.

10:48But if the agent is doing that, that's a much higher return to scale if that human is sort of out of the loop. Now, the human can be on top of the loop, so to speak. And on top of the loop means they're engaging in an imperative. Go do this. And they're not sitting in the middle engaging in the interrogatives. interrogatives. Constituent agents are engaging in the interrogatives. And that's what I mean by there's a Socraticism that happens between the command to go do something and the final output. And this is really interesting because, I mean, it still maintains some of the, I'd say, black box concerns folks have at the model level.

11:31But because of the way you're constructing these agents and how they are approaching things, it sounds like there's a much better auditing process for what's actually happening there that then the human employees can focus on okay let's audit to make sure this is working and that we're actually delivering on these processes whereas i'll say currently one of the big things we noticed in that bot report i mentioned was that 96 of all basic bot created prs that are being sent out right now are not linked to product management tools well if i have an agent doing that i presume we can simply make sure the agent is using a project management tool to track this and kind of improve the observability of a lot of this code that's coming into the code base and a lot of the approaches we're taking.

12:17At the present moment, we're able to trace the steps that an agent takes to complete tasks. And in fact, we're able to hash and sort of have an auditable record of the reason process. and so it's not at the workflow level it's not fully black box now one could argue that again these agents have access to llms and the llms have sort of deep learning models and maybe there's some black box models at the root of sort of the language models that's that's definitely i mean that's probably the case i mean when i think about older so like the last decade where we're building machine learning models. Some of those non-parametric models, they're black box.

13:07It's really hard to interpret what's actually going on. And that's different from the parametric models where you have a prior assumption as to the distributions of the fields that comprise the training set. And I'm not sure in all cases, the black box nature of those models was necessarily an issue. I hear you. It's interesting, though, because I think a lot of survey data will show that the concerns that when you talk to a CIO, a CISO, or just a general software engineer is they're saying, hey, I don't get what's happening here. I mean, we can also talk about the concerns of is this going to replace me?

13:49I think we will get to that in the labor economics side of this. But what would you say to those folks who are just worried about implementing this and not being able to unpack when something goes wrong or other concerns on that kind of black box piece? Well, I think if we're talking about agents and agentic frameworks, then a lot of the reasoning is there's layers of reasoning that are abstracted above and beyond the deep learning models that are sort of at the base of these LLMs, that set of reasoning processes, you're able to cache that, you're able to trace it, and you're able to audit it.

14:33So at that level, I think, and generally, as I think about this, the worker, maybe the software engineer or whatnot, they would want access to that layer of reasoning. I'm not sure they necessarily need access to reasoning that's happening, so to speak, or the inferencing that's happening, so to speak, at the LLM layer. Do you think software engineering is going to be one of the disciplines that particularly benefits from this increasing use of agentic AI and its ability to autonomously handle tasks? I think so. I think that software engineering is going to benefit as other disciplines, knowledge work disciplines.

15:15I think it's going to benefit from sort of an industrial engineering paradigm, right? So if you go back in time, you have, let's say, factories with seamstresses who are designing, fabricating, sewing effectively clothing. But the industrial revolution came along and said, okay, we're going to design the factories and the systems that output those articles of clothing. And software engineers similarly work. In fact, arguably, in some regard, the work of an AI engineer is really to kind of build those systems that effectively build systems. I guess if you're building agentic frameworks and then deploying those agents to effectively write code, and then the code is for the purposes of building an application, then you're building the systems that effectively build the constituent systems that build the application.

16:12And in that regard, it's sort of more like industrial engineering than it is necessarily software engineering. Or I would almost describe it as more of an architect mindset versus a carpenter mindset. that. And we've already talked about that kind of, I mean, plenty of people have an architect title already, and I think we're going to see a push more towards that. What are the other industries that you expect to see a lot of changes coming due to agentic AI improvements? Well, I think in terms of roles within the knowledge work, the arena of knowledge work, I envision these agents climbing the career ladder.

16:47So initially they can do some data entry, Then they can book meetings. Then they can send emails out and monitor maybe a drip campaign. And then eventually they just go up this career ladder. They progress. Maybe they write some basic applications. I do think we're far from the agents necessarily solving for product market fit or doing more designing products. I need to keep my job. I think sort of the more complex and creative job functions, as you say, for example, architecture, it's not necessarily, that's a layer abstracted away from doing specific tasks. So in that regard, I see a whole myriad of especially task-based work being disrupted.

17:35Something I'm realizing when I interact with companies is that there seems to be a set of three or four stages prior to even engaging, but each of those stages, the customers are willing to pay for. And the first is enterprises, they want people to explain to them how AI fits into Evergent. They don't have, let's say, necessarily an immediate answer to the use case within the organization that can benefit from AI. And I actually think that type of reasoning is this design of experiment reasoning. and you learn about that in human factors engineering you learn about it in research methods certainly if you're doing machine learning work you have to sort of say okay what is the use case but you're reducing that use case down to a single variable called the target variable the dependent variable and you're aligning a bunch of data that you have to sort of bear results for that target variable and that's a that's a way of thinking about the world.

18:43And that is sort of like an experimental form of reasoning. And that lends itself to kind of going into an organization, figuring out how things are functioning, and then figuring out how AI can benefit those systems and processes. And right now, I noticed that companies just want help with that. And some subset of those companies, I think as a sort of an emerging AI startup, those are your pilots and then eventually those are your msas but i think bain capital got a report recently that i thought was interesting showed some hurdles that have been overcome insofar as ai adoption some of those hurdles were infosec related you know keeping private data private but the other hurdles that have yet to be overcome sort of like what's the use case It's like kind of a classic thing.

19:38What's the use case? How do we apply this? How do we wrap our minds around this? It seems like the role of the AI engineer right now is really that. And then eventually that translates into ingesting data, ingesting content, building systems, analyzing the output. Based on the output, where's the refinement? And then based on all that, can you get agents to kind of automate processes? but there's a lot of work right now to be done just informationally so the work will stop there by the way and and it really shouldn't just stop there that's not that's not the end all deal but it's um it's certainly a it's certainly a need right now it feels like part of what's happening is we know there's so much potential to apply ai and there's so many opportunities that zeroing in on where to start and kind of the steps to take along the path is a challenge for a lot of folks right now.

20:32Precisely. One way I would think about this design and experiment reasoning is that you've identified the why and you can distill that through the what into the how. There's a lot of content out there that, okay, the why is getting to the point of being self-explanatory. How do you distill that down to the how? In the short run, you've noted that AI will encourage engineers to think more Socratically, to ask more questions, to really elicit more. What does that mean for day-to-day development now? And then I'm curious where you think it's going to go in the future. I don't know if I necessarily think that in the short run, the agents are going to promote increased Socratism by engineers specifically.

21:18I don't know. I mean, I think more specifically, the agents themselves will take on that socratic reason and so you think the shift is already happening to a more managerial mindset for engineers where it's like hey i'm going to manage these agents i need to make sure they're they're functioning well versus let me elicit info from i think that the shift is happening where the human being commands the agents that's the imperatives. And then the reasoning that needs to happen between the imperative and the completed task, no pun intended, begs a lot of questions. And it's that interrogative process that I sort of think of as like the Socratic dialogue.

22:03Got it. And then obviously on the more I'll call it individual level, we're all engaging in that Socratic dialogue when we simply use ChatGPT to like rewrite an email for us. Precisely. And when we're engaging in that sympathetic dialogue, we're maintaining context in our minds to sort of guide the conversation towards an outcome, whether it's winnowing down the reasoning process towards something that we want with increased specificity. That process, again, depends on the access to long-term memory, the access to short-term memory, for adaptability, training these agents on particular tasks. Again, there's a difference between knowledge and expertise.

22:43So there's a knowledge layer here, with these agents and there's sort of like an expertise layer. And for example, you see that, for example, if you want to have a contract written, you can leverage chat GPT to edit various clauses and this, but you'd probably want someone with expertise to just look it over, see if it is cogent. That is sort of where a lot of this modeling context comes into play. And all of that modeling context, as I see it, is a bus between sort of the foundational layer and the application layer. And it's really all of us right now with regard to these agentic frameworks, we're sort of just developing the bus that sits between the CPU analogy and the application layer analogy.

23:31Like I think many folks listening, I am only at this point using kind of the Socratic approach. I'm leveraging maybe Copilot. Maybe I'm just leveraging ChatGPT to help me with certain test writing tasks. I'm not yet. leveraging agents, but I see that it's going that way and I want to start leveraging agents within my software engineering team. What would be your advice to those folks who are listening and saying, damn, okay, here's the next level. I need to start moving there. One advice is to say, with regard to the labor market transaction, there's two sides. And there's a sell side of the labor market transaction.

24:06There's the buy side. It will be, I think, increasingly difficult to compete with agents on the sell side of that labor market transaction. So if you're on the buy side of that transaction, then these agents are great because they've reduced your operational expenditure, etc. So the way I see it is that we as humans will probably leverage these agents to do work. And if we can own the relational structure, we own the product of the labor and would kind of own the relational structure, then the humans directly benefit from the agents. But if you don't have ownership over that relational structure and you don't have ownership over that work product, then effectively you're competing with the agents.

24:55Yeah, and I think this is where a lot of the fear currently comes in for folks, where they hear that and they go, wait, does that mean I need to run my own company? Does that mean I need to go be an entrepreneur? does that mean I'm going to not be as strong with the labor market in a couple of years as I am today? Like, what do I need to do to change that? And obviously, AI has extremely broad implications on what's happening with the labor economy. How do you see AI agents impacting the broader structure of work and labor contracts in the coming years? Well, I think the roles-based economy is too rigid.

25:30So I see that companies have roles. And again, like I said, those roles mapped to jobs, to workloads, to tasks, et cetera. And roles offer a degree of economies of scale for these companies. Agents actually, by the way, offer economies of scope. Because the way to think about an agent is you're designing a tooling line that can effectively manufacture various products. To your comparison earlier, I'm building the factory versus I'm building the digital piece them a lot. Right, right. I generally think that if you're building agents into the enterprise, and that inclusive of workflows, but also sort of culturally and structurally, then you may want to rethink this sort of concept of a role.

26:23Certainly when you're starting your own company, when you're functioning as an entrepreneur, you're not engaging day in, day out at the task level or at a single workflow level. You don't just have this one job that maps to this one role. In my opinion, that's very junior. Another way of thinking about it is the roles can be assigned to these agents. If the roles are being imparted to agents, How does that redefine the roles-based labor economy? And if, again, you're going back and saying these effective workers who can sort of think Socratically, in many cases, they're polymaths. And if they can go in between various functions and the agents are sort of doing the work, then why should they sit and just do one role all day long?

27:19This doesn't make any sense whatsoever. So all of that, I think, if you're thinking about updating the frameworks of the labor economy, I think you slowly want to start thinking about the rigidity of the roles-based labor market. I mean, we're not all sort of sitting on a factory line and sort of at that level in sort of the knowledge work economy. And yet a lot of how we've built work is based off of that factory model, to your point. Precisely. Which is very early 20th century. Fascinating. I wonder if this is going to exacerbate or kind of remove the trend of like freelance work, of gig work that we've seen increasingly over the last 10 years.

28:04Do you think that agents are really going to replace freelancers or are you going to see this kind of freelancer gig work approach where, hey, I'm running agents on behalf of a company now? because I know, and you've talked about this before, that many companies, many large enterprises are taking this insourcing approach and saying, well, if I can build an agent to solve these tasks for me, I'd rather be able to have broader observability and potentially cost savings compared to, say, passing this basic data task off to another country where it's cheaper. I think there's a, even, let me go one step more broadly.

28:41I think there's a case to be made for vertical integration. I think that these agentic frameworks inherently have a multi-layer stack. You have the foundational layer. You have the long-term memory layer. You have the short-term memory for adaptability. You have these agents that are trained on tasks. You have this integration layer. And there are feedback loops that, in fact, there's a two-fold feedback loop that just comes to mind right now. And that is to say, if you're allocating these agents to specific workflows within an organization and they're improving processes, okay, there's an improvement in so far as the returns to scale in so far as those process improvements.

29:27But then also the outcome of those process improvements and the knowledge derived therein can go back and train the models and effectively train the agents to kind of further that process. And it's very, very difficult to have those feedback processes without vertical integration. So that's one point. There's a second point here, which is that if you think of this multi-layered stack, if you will, again, and go back, let's say 10 years and think about an analogy to AI, a very good analogy to AI is sort of like machine learning set of companies in the early 2010s. And those companies were, the ML models were like an afterthought to the use case and the application.

30:15And so you're sort of building at the application layer and then, yeah, you may have an ML model somewhere and that ML model is benefiting the application, but you're really thinking at a top-down perspective. With these AI systems, everything's really happening bottoms up. The foundational layer comes about first, then the RAG layer, then the short-term memory layer, then these agentic frameworks. And frankly speaking, the application layer, the AI application layer, sort of maybe still around the corner. So if all that stuff is sort of happening bottoms up, well, then that begs two interesting questions.