Planview Chief Data Scientist Dr. Rich Sonnenblick on How AI Agents Will Transform Software Development and Use

Hey guys, thanks with Throw, we're here with Richard Sonin Blick, who is Chief data Scientist for planview. And we're talking about how AI and AI agents specifically are transforming the way we think about software development and also the implications that has for governance in the enterprise. Richard, welcome to the show.

Thank you so much, Mike. It's a pleasure to be here. Everybody and his brother is talking about AI agents, but I'm not sure we fully have thought through the implications.

Um, in theory, we're not only gonna build AI agents, but then the AI agents are gonna build applications for us, and then we're gonna have to figure out maybe how to govern all that stuff and apply some controls to that. Give us the big picture here. Where are we on this journey?

Yeah, that's a great question. And, um, you know, we've all been familiar with large language models and using chat GPT using Claude to answer questions. And it's, it's been amazing to see what, uh, large language models can accomplish on their own, but at the same time, um, they're really just chat bots until you add this agenta capability.

And so with agents for, um, US at Planview, we're talking about two things that, um, first involve reasoning and the ability to break down a problem into a set of steps and then to, um, do tool use or perform tool use and act on each of those steps using some set of capabilities that you've given the LLM. And so it's this reasoning and acting working together that give, um, l LMS the ability to be agents, right? And, and that reasoning and agentic capability, uh, then, then allows the agent to go do things in the real world, whether that's, uh, do some shopping on your behalf or, um, build a database for a development project, um, or, uh, perform, uh, some sort of project management activity looking for work that is approaching its start date but not yet ready to actually be commenced.

Um, those are all jobs for agents. And the other thing that agents do that is critical is they can talk to other agents. So that agent that might be looking at, um, uh, booking a reservation for you is going to talk to another agent that has access to your calendar and make sure that the days that it's, um, uh, setting up travel for you are days that you're free and are the days that you're planning to travel That agent that's looking at project management activities that are coming up but not yet ready could talk to another agent that has access to the team members and can slack them and say, Hey, this work is almost ready to start, but we haven't finished filling out the remit of what the definition, the scope, and the activities are.

Can you help with that? And so it's this ability for agents to talk to each other that takes agents beyond just the ability of, uh, breaking some problem down into steps and, and performing those steps one after the other. Hmm.

Well, the great thing about it, as you described it is AI agents are autonomous. The scary thing about AI agents is they're autonomous. So how do I kind of put something around this that prevents them from doing things that previously we would've had some sort of control for IEI don't give developers access to all the data in the world, and yet if I look at AI agents, it seems like they can get access to data in ways that I never imagined.

So how do we kind of think this through a little bit so that the mm-hmm. Proverbial cure doesn't wind up being worse than the disease? Yeah, that's such a good question.

Um, governance looms large around agents, right? And at, at the 35,000 foot level, if I have an agent that I am imbuing with some level of authority, um, at the very least, we need to make sure that that agent can't do more than I can do. So it, it can, um, access the repos and GitHub that I don't have access to.

I can't write to the repos that I can't write to. Um, but beyond that, there's probably, uh, imperative to even reduce the scope further. So thinking about what this agent does, maybe it only has access to one repo or within my resource planning solutions, it only has access to, um, one of the Kanban boards or one of the, uh, projects or initiatives that I have access to.

And maybe it only has read access or maybe it also has right access, but that right access is predicated on an approval that it's going to come to me first, um, and maybe send me a Slack message, maybe send me an email, maybe give me a toaster notification within a planview application that it wants to do something on my behalf. Here's what it's going to do, here's why it's going to do it. So there's transparency and, um, explainability.

Um, and yes or no, do I authorize that behavior to occur? Right? Um, and then taking it one step further and thinking about the ability to then review in an audit trail fashion, all of the things that that agent has done on my behalf.

Or if I have a team of, um, of developers, all of the agents that they have asked to do things on their behalf and what those things were when they authorized them, so that there's that complete audit trail there as well. And then perhaps even when possible, the ability to undo that. So I can see what my agents have done, I can see that, yeah, I did authorize these things or I authorize something to happen automatically without them even coming back and asking me for an approval, but now I can roll that back 'cause I've thought better of it, didn't really understand the implications of what the aging was asking me to authorize.

Let me ask you, I think there's generally a lot of excitement about AI agents, and yet we don't really see them being rolled out in masks just yet. And I can't help but wonder if it's not because of the issues we're discussing here. The the devil is in the details and people are trying to sort through all these things, and they're more complicated than most people realize.

I, I think it is, um, more complicated than what people realize. And I think it's difficult for us to imagine what agents are pos uh, what agents are capable of and what's possible, um, and what's helpful. And, and I think when, when I think about, um, the art of the possible with agents, um, I go back to my own product manager training and think about personas, um, whether it's, you know, a, a developer persona, a database administrator persona, or, um, a, uh, product manager persona or a CTO persona, and what are the jobs to be done associated with each of those personas and which of those jobs should be or could be done with an agent, right?

So, so it's easy for, I I think most people to think about, well, I'm going to have an agent that, um, runs out and looks at a bunch of stocks that I'm interested in and comes back to me every time one of these stocks says, say, changed in value every, you know, by 5% or more. But it's much more challenging to think about, um, I'm a product manager. What are the agents that I should be building that can replace things that I do that, um, that will actually add value to my job?

And, and so I, I think as, as software developers, um, plan view, we have an obligation not just to give an agent construction set to our customers and say, here, build your own agents, which we do in the context of planview copilot, but also to provide out of the box agents that we know align well with the jobs to be done that, um, the personas embodied by our customers and our users are working through every day. Mm-hmm. Do you think this will change the traditional build versus buy conversation that a lot of enterprises have?

And I'm asking the question because it's always been more challenging to build something. So I went and I bought a packaged application to go do something, and of course, that package application will now come with AI agents. However, if the cost of creating an AI agent or, or a piece of software drops dramatically, will more people shift towards building something that is custom to their particular use case?

Right? I, I think as, um, as the agents that are out there gain capabilities, we'll see more and more users or companies building agents that are separate from the existing enterprise software packages. So building agents that can go directly into, um, uh, your, your cloud solutions may be SAP or Oracle and do things on your behalf, but those agents were not written by SAP or Oracle.

So you may be using an agent construction set by one of the many startups that are, um, providing that capability today rather than say, going to your SAP um, uh, team and saying, Hey, we wanna add agent capability. How much is it gonna cost to get your SAP agents? Um, so, so right now what we're seeing at Planview is that our customers are looking to us to provide both the agent construction capability and out of the box agents, but I can see a future not too distant from now where, um, people will have their agent builder capability that's sitting outside all of their existing enterprise packages and they're using those builders to do things that are across many of the different applications that they use on a daily basis.

And that's actually where we are heading at Planview with our data foundation, which pulls in information not just from Planview applications, but also perhaps ServiceNow or SAP or Atlassian applications. Anywhere you have a system of record that involves work, the resources, the strategic objectives that you're working on, um, so that we can provide AI agents on top of all of those systems of record that provide insights not just on the work that may be in progress within planview, but within all of those systems. And I think that kind of trend of coalescing information across these systems of record is only going to accelerate with the, um, uh, with the increase in uptake that we're seeing in agents.

Earlier you were discussing identifying things that I have to do and assigning that to the agent. And a lot of that comes down to us doing things the way we've always done them historically and the way we go. But I feel like every time there's an innovation, we do the same thing over and over again.

We use it to do the same thing we always do a little bit faster, but are we gonna get to the point now where maybe we can just reinvent what it is we're trying to do and maybe not do it the same old way and and just blow it all up and re-engineer it end to end? Yeah. So that's an interesting idea.

And can we, um, can we use LLMs, if I understand what you're saying, Mike, can we use LLMs to build an application on the fly? Um, that might take, and, and, and that application might completely reinvent something. I mean, I'll give you an example.

We're so siloed today and between sales and marketing that we have all these different teams and yet they are kind of natural extensions of each other. So maybe AI agents will change the way we think about sales and marketing, for example. Mm-hmm.

I, I think that that's going to happen and I think before that happens, it's going to, uh, agents will uplevel all sales and marketing, um, teams to be the best at what they do, right? So if, if we have the agents embodying the best practices of what it looks like to be a proactive product manager or a proactive product marketing specialist, um, and those agents are doing things like doing the external landscape assessments, um, and updating those for us on a weekly basis, um, then we are actually, um, skilling up as we go, right? The product marketing that I do is going to be improved because I don't have to remember to update those landscapes or, or wait till my boss tells me to update those landscapes.

It's being done on my behalf and being brought to me. Um, and so I think that first is going to revolutionize the way product teams work with their development organizations, product teams, work with their marketing organizations, because the agents are going to be doing all of the connective tissue work that we might forget to do on our behalf and simply bringing us the, um, the out outcomes of those conversations, the outcomes of those insights, and, and then even suggesting the next set of actions that we should take. And then I think what you're saying a hundred percent, we are going to see novel ways of say product marketing, product engineering, and uh, uh, product management organizations working together because the information flows are going to be so much more, um, ever present.

And the amount of data coming across these, um, different pieces of the organization are going to be, um, it's gonna be so much more high bandwidth. So it is going to dramatically change the way these, these teams work that may have been very siloed in the past. Um, you hear people say, and I'm not sure if it's hyperbole or not, so I'm gonna put it to you, then we might be building more software in the next two years than we developed and deployed in the last decade.

Is that feasible in your mind? Or, and, and are those real applications, or are they just kinda like these little personal AI widgets that I have on the side and we're calling those apps, but you know, what's your expectation here? Yeah.

Well, if you just look at velocity, uh, you know, hands on keyboard or agents on keyboard, um, certainly the number of lines of code is increasing dramatically. Um, but we can't measure the number of deployed packages by lines of code alone, right? So I think a few different things are happening.

When you think about the gestation of a new application, the, um, the barriers to that are coming down, right? The ability to vibe code, whether you are a seasoned developer or you are someone who's done marketing all your life or you're a school teacher, the ability to vibe code is dramatically changing. The, um, uh, who, who gets, who gets to benefit from a simple application.

And it's also changing the rules of the game for startups where you don't need, uh, you know, a head engineer and 10 developers to build a startup. You can build a startup to some level of maturity with, you know, three people, two people. There are startups out there that are, are getting a million or 2 million, uh, dollars in annualized sales today are just one person by coding and then making adjustments to that code themselves and getting it out in front of customers.

So, so there's a, there's a lot that can happen there. Is it, is it dramatically changing the number of applications that software organizations are selling? No, but I think it's changing who uses software?

So going back to that school school teacher example, if you think about, um, uh, say a, a school teacher who, who needs to analyze the test scores across a cohort of students, um, in the past they might have done that with a spreadsheet, but now they can vibe code something and use that application, uh, week in, week out, um, term after term. And they are, you know, becoming in a way a software developer as they do that. And I think as people realize that they can be sort of software developers through VOD coding, we're going to see everyone building, uh, small applications that are maybe too bespoke or too specific to actually be something that someone else creates and then sells.

But, um, you know, we'll see, we'll see people building these things themselves and using them privately. And I think that explosion that is already happening, um, as I talk to people who have never coded before and who are, you know, working on documentation, uh, who are, uh, professors of history or professors of anthropology, uh, these are people who never thought they'd build anything like a script or a program, and they're now doing that and, and they're saving tons of time through workflows that they never saw through the lens of programming. Do you think that our backend workflows are set up for this?

And I asked this question because we're already run into all kinds of issues where the pipelines are too brittle, we don't have enough pipelines. The project management app is kind of clunky and it doesn't connect to the communications platform, whether it's Slack or whatever else. And, um, you know, sometimes I feel like we build applications in spite of our tools, not because of them.

Um, will we go back and revisit all of that and say, geez, you know, in a world where everybody is a developer, we need to rethink what that entire workflow looks like? Yeah, that's a good question. Um, we are going to need to rethink it and, and I think, um, software organizations that put agents in front of school teachers, um, that that can develop, uh, an application also need to put agents in front of those same people that can test those applications and can anticipate how to validate those applications on the user's behalf.

So again, just like we were talking about the jobs to be done, um, being embodied within the agents or reflected in the agents, um, if we can have the software development lifecycle reflected in a set of agents so that people who don't know what software development is really all about still benefit from the best practices around software, whether that's code coverage or, you know, unit tests or stress tests so that they're not getting in their own way, um, as they vibe code solutions. And there's obviously a lot we can do to make the whole process of AgTech software development more resilient. And, and I think we're seeing the baby steps in that direction with Claude code, with, um, Gemini, CLI and all of the other agentic systems that are out there.

They're adding these pieces so that you don't even need to know that those pieces should be part of the process. They're just inserted on your behalf and, and helping make what you build, whether it's by coding or something more hands-on, more resilient. All right.

Well folks, I think we've made it clear that AI agents are a cause Now we're just waiting to see what the effect is gonna be. Hey Richard, thanks for being on the show. My pleasure, Mike.

Thanks for having me. Alright, and back to you guys and Steve.