An Hour of Innovation
An Hour of Innovation
This podcast explores the art and science of innovation through conversations with product, tech, and AI leaders, scientists, and innovators.
Oct. 6, 2026

When AI Can Build Features, Who Decides What’s Worth Building? | Tammuz Dubnov

When AI Can Build Features, Who Decides What’s Worth Building? | Tammuz Dubnov
An Hour of Innovation
When AI Can Build Features, Who Decides What’s Worth Building? | Tammuz Dubnov

In this episode of An Hour of Innovation podcast, Vit Lyoshin speaks with Tammuz Dubnov, co-founder and CTO of AutonomyAI, about what happens when AI coding agents move from impressive demos into real production products with legacy code, dependencies, architecture, testing, review queues, and product judgment.

Vit and Tammuz discuss why building a small app with AI is very different from changing an existing codebase. Tammuz explains how AI-generated pull requests can look finished to a product manager but still be too large, unmaintainable, or hard for engineers to merge. The conversation separates demo magic from production reality: a feature is not truly done just because it appears to work.

They explore why the harness around an AI agent matters as much as the model itself. Tammuz describes how a product-focused agent needs to translate intent, find the right files, avoid irrelevant code changes, run checks, test the product visually, and verify that the result matches the user experience the team wanted.

Vit and Tammuz also cover how AutonomyAI helps product teams turn ideas into briefs, sketches, tickets, presentations, prototypes, and pull requests. They discuss why code is the source of truth for product work, how PMs can use agents without staring at code, and where engineers should still own sensitive technical decisions such as architecture, schemas, backend logic, and merge quality.

The episode closes with a practical look at autonomous product delivery, product judgment, and the new bottlenecks created by faster building. Tammuz argues that agents can help with analytics gaps, accessibility, friction, bugs, and peripheral product improvements, but PMs remain the gatekeepers. When AI can build more features, teams still need humans to decide what should ship, what deserves review, and how to avoid overwhelming engineering with work that should never reach production.

Tammuz Dubnov is the CTO and co-founder of AutonomyAI, a platform helping product teams turn intent into delivery inside production-grade codebases. His work focuses on AI agents, product development workflows, autonomous product delivery, and helping product teams move from ideas to working software while keeping the right human judgment and engineering safeguards in place.

You’ll learn

  • Why building a small AI-generated app is different from changing a legacy production product
  • How AI-generated PRs can look functional but fail engineering review
  • Why the harness around an AI coding agent matters as much as the model
  • Whether vibe coding will remain a skill or disappear into better platforms
  • How AutonomyAI helps product teams turn intent into tickets, sketches, presentations, and pull requests
  • Why code is the source of truth for product work
  • How agents can use analytics, customer signals, and team knowledge to suggest product improvements
  • Where PMs should have authority and where engineers still need to own technical decisions
  • How autonomous product delivery can maintain analytics, accessibility, bugs, and product friction
  • Why PMs remain responsible for deciding what belongs in the product
  • How faster feature creation changes prioritization, review queues, and product judgment
  • What PMs and engineers can try now if they want to use AI agents more effectively

Timestamps

00:00 Introduction

01:16 AI in Production Codebases

03:01 What Impressive AI-Built Features Look Like

05:49 Why Agent Harnesses Matter

08:22 The Future Beyond Vibe Coding

10:29 Autonomy AI: From Intent to Delivery

14:06 A PM Workflow: Idea to Pull Request

17:16 Turning Team Knowledge into Product Signals

22:18 Guardrails for Safe Product Changes

26:42 Faster Iteration and Team Collaboration

30:35 Autonomous Product Delivery

34:32 Human Judgment and PM Skills

39:13 Prioritizing in a Faster Building Cycle

42:41 Where Teams Should Start with Agents

44:22 The Product Team of the Future

46:36 Innovation Q&A

Connect with Tammuz

Connect with Vit

Support the Podcast

If you enjoy the podcast, you can support it by exploring the tools below.

These are affiliate links, meaning the show earns a small commission at no extra cost to you.

  • Kit: Email marketing that automates your growth.
  • Google Workspace: A collaborative way of working in the cloud, from anywhere, on any device.
  • Riverside: AI-powered platform that lets you record, edit, repurpose, and distribute studio-quality content as easily as if you had a crew behind you.

For inquiries about sponsoring An Hour of Innovation, email iris@anhourofinnovation.com

Episode references & glossary

Claude Code
https://claude.ai/code
Claude Code is mentioned in the episode materials as an example of a coding agent a team may already use.

OpenAI Codex
https://openai.com/codex/
OpenAI Codex is mentioned in the episode materials as another example of an AI coding agent used for engineering work.

Lovable
https://lovable.dev/
Lovable is referenced when Tammuz compares AutonomyAI’s interface pattern to a chat-plus-preview AI building workflow.

ChatGPT
https://chatgpt.com/
ChatGPT is mentioned when Tammuz discusses how people currently use individual threads to explore feature ideas.

Mixpanel
https://mixpanel.com/
Mixpanel is mentioned as an example analytics connector agents can use to understand user behavior and shape product plans.

Figma
https://www.figma.com/
Figma is mentioned as a design abstraction that may describe intent but is not the same as the actual product behavior in code.

Slack
https://slack.com/
Slack is mentioned as a collaboration channel where agent tasks and product work can be shared with teammates.

Three.js
https://threejs.org/
Three.js is mentioned as an example of impressive visual demo work that may not necessarily apply to production product environments.

Blender
https://www.blender.org/
Blender is mentioned as another example of impressive AI demo output that Tammuz distinguishes from production product work.

Vit Lyoshin (00:00.86)
Welcome Tamuuz, thank you for your time today.

Tammuz Dubnov (00:03.736)
Thank you for having me. Very excited to be here.

Vit Lyoshin (00:06.864)
Yeah. So let's start with this question. a lot of people can make an application these days with AI, but beyond the like demo and personal projects, what have you seen people building for like real production environments?

Tammuz Dubnov (00:26.668)
Yeah. so the game is between are you making something brand new, which is small, easy, also context-wise, context size for LMs, two are you working in something that that's legacy often huge? those products are are much more complicated to work within. There's a lot of dependencies, a lot of complications. and I see that people are maybe able to create kind of standalone small apps, but as soon as you jump to contributing to real existing code base of

Vit Lyoshin (00:40.668)
Mm-hmm.

Tammuz Dubnov (00:56.96)
A product that can be again massive in legacy. That's where you see that kind of dropping off. I've seen people open PRs, huge PRs, unintentionally so. PM saying, hey, look at this feature I built, everything works. my God, let's have it out. To engineers looking at the PRs saying, this is five times as long as it should be. The code is not maintainable. I'm not gonna bring this in. The code doesn't reuse. Like there's all these technical requirements.

that have really solid foundations to why they're necessary. but if you're you're non-developer, you don't always see it. If you kind of vibe code your way through. And so I I have seen partly frankly through us PMs that do do it successfully and open PRs that are frankly small and target exactly the part of the code and the feature that they want to change and and do actually go smoothly. but if you go about it vibe coding like you would an app, it it doesn't hold through.

Vit Lyoshin (01:47.666)
Mm-hmm.

Vit Lyoshin (01:52.08)
Yeah, yeah. Have you th seen anything built with AI that like really surprised you? Like how awesome it was?

Tammuz Dubnov (02:01.07)
I'm not gonna jump to like the the demos because the the demos are typically really small apps again. And then yeah, you could do wow factors like 3js and and blender and and beautiful stuff people are doing, but this doesn't actually apply to production environments. for me in the production side helping examples from from our side, from stuff our RPM, for example, has done where she's taken like a feature end to end stuff that

Vit Lyoshin (02:05.616)
Sure.

Vit Lyoshin (02:16.721)
Mm.

Tammuz Dubnov (02:29.986)
We would have thought in the past we'd need engineers. she surprises us and and just gives us everything working, including code that is solid and and we can effectively merge. so I think that surprised like she did one really nice feature. our UX is is kind of similar to Lovable in a sense. You chat, and on the right, there's always the the part of the app you're working on. And she added voice mode to everything. so you can record. There's like a beautiful animation of

the your sound waves as you're recording and it shows you transcribing. And she added a also an option to text to speech where it reads it aloud. and just a bunch of stuff that even like models involved for transcribing and reading aloud. And she just did all of it. and it's super solid and frankly it looks beautiful. So like that feature was was nice and impressive on her end for sure.

Vit Lyoshin (03:27.259)
Yeah, I'm interested. I wonder if this is also specific to a person like knowing exactly like this at the skill level, right? Knowing exactly how well describe it so agent understands and builds exactly what she was asking for. or is it also partially like on the technical side, the quality of the model itself and how well it's like optimized for certain tasks? that's

That's something I asked myself when I was trying to use wipe coding and and figure it out. Sometimes it works nicely, sometimes it doesn't. And I always wonder which one is it?

Tammuz Dubnov (04:05.966)
I I would say it it's a mix, but nowadays all the models are pretty amazing to the extent that it falls mostly on the harness. like in autonomy, that a lot of our focus is on the product harness we built. so it's the harness's job to translate the the little bit of you what you gave it in terms of your intent into what is it gonna change? How is it gonna verify that the agent did the work? How is it gonna prevent the agent from touching 30 files that are totally not relevant to what you really asked for? so

Vit Lyoshin (04:13.466)
Mm-hmm.

Mm-hmm.

Vit Lyoshin (04:23.76)
Mm-hmm.

Vit Lyoshin (04:34.352)
Uh-huh.

Tammuz Dubnov (04:35.83)
Models are are amazing. The harness is what keeps them in line.

Vit Lyoshin (04:40.035)
Yeah, can you explain a little bit more this hardness s for for people? This your concept?

Tammuz Dubnov (04:45.036)
Yeah. yeah, absolutely. So let's take a second to think how you interface with with all these, like vibe coding like interfaces. you give it a a short prompt that effectively that prompt is is not the work it's gonna do, it's a goal it's meant to reach. you're saying add this button, but it's it doesn't just add a button. It needs to go and like I don't know, run the ad, figure out what button components you have, figure on the page where it needs to position it, how it reacts and how it's responsive, and then what it triggers.

Vit Lyoshin (04:54.576)
Mm-hmm.

Tammuz Dubnov (05:13.878)
So you just gave it the goal and now it needs to do a bunch of work. It needs to read a bunch of files. It needs to edit. It needs to run and check itself. Maybe you have testing, it needs to run its testing suite. And the harness helps it stay on track. The the struggle is you wrote your little bit of instructions, the goal. That can get lost in a ton of context. It has to accumulate in the project. Like it needs to read, I don't know, 30 files to find the two files that are relevant. Well, how do you keep the other 28 files it read that aren't relevant?

Vit Lyoshin (05:39.056)
Mm-hmm.

Tammuz Dubnov (05:43.289)
How do keep them from distracting the agent? Because the agent is playing a game of distraction of, my God, my context has all this stuff, but I don't know. 80% is not relevant to the work I need to do. And that's where the harness comes in. It keeps the agent in check. It checks if the agent's kind of stepping out of bounds. It brings it back. If the agent said it finished the work, well, the harness can check, did it finish the work? If not, it pings and say, no, you didn't actually finish it. Or yes, you finished it, but this test is failing. Go back and check your work.

Vit Lyoshin (05:45.583)
Mm-hmm.

Tammuz Dubnov (06:11.85)
Our harness is really focused again on products. So even if the agent says it's done, it says amazing. Now agent run it and interact with the feature and see with screenshots and like actual browser interaction that what you think you built is what you built. Because what it's changes in the code is not really how it comes out in the product. You have it to do that testing. And the harness forces it to do all of that. So it does a proper workflow. And again, all you need to do is give that little prompt you had in the beginning. That is effectively the end goal.

Vit Lyoshin (06:14.991)
Mm-hmm.

Vit Lyoshin (06:25.583)
Mm-hmm.

Vit Lyoshin (06:39.415)
Mm-hmm. Yeah, okay. So starting from the end, you wanna have descri a description of what your end result should look like. And then it goes, checks what needs to be built, and then actually builds it, verifies it, and compares it with with that call. That okay, that makes sense. Interesting.

Tammuz Dubnov (06:59.116)
Yeah. I I will say some people also like to work iteratively. So they give like a a general goal and now that they're there, they see, okay, well, how do I like the shape and feel? And now I'm gonna move you a little bit to the left, move you a little bit to the right. That's also a very healthy way to work with it.

Vit Lyoshin (07:02.693)
Uh-huh.

Vit Lyoshin (07:10.619)
Mm-hmm.

Yeah, yeah, okay. So and also about wipe coding, do you think it's gonna be become like its own job, I guess, or its own skill that people will use once in a while or in a certain cases, or is it gonna become more normal for product teams to utilize, for example, like pretty much everybody use email, for example, or like Excel?

in finance or right if you financial professional, you use Excel, you know Excel, like that's your tool. what what do you think about this vibe coding? Where is it going?

Tammuz Dubnov (07:51.024)
potentially controversial, but I think it's gonna die out. I don't think it's gonna stick. in the sense that you used to have like the role of of prompt engineer. Now you you don't really hear that as much. Why? Because everybody can can prompt. And is it did everybody go through that education to be what it once required to be prompt engineer? No. The models, the the harness, the platforms, they got so so good.

Vit Lyoshin (08:03.503)
Mm-hmm.

Tammuz Dubnov (08:17.259)
You don't need to engineer them anymore. You can just prompt them like you would, and they do the engineering and they manage. And I think the same thing is gonna come with vibe coding. there's engineers using it, but typically when we're talking about vibe coding, we're talking about the non-engineers. And if you're a non-engineer, then why do you need the coding part? You don't really look at the code. I think a platform is gonna encapsulate well, we do it also in what we build, but encapsulate it entirely.

You're gonna get a no-code experience. You don't need to think about any of that abstractions that developers need to worry about, like even Git. I think the where it's going is you'll just be a builder. You don't need to think about vibe coding, because everything to do with coding is gonna be taken care of by the platform. And then you can just go and build. Say what you want, the system will do it, it'll show you, you'll interact, and all the the hard work people are putting in now.

to vibe code and see that the code that it generates works, then that's gonna go. Some that's gonna be abstracted away as everything keeps getting better and it just it's taken care of for you.

Vit Lyoshin (09:15.089)
Mm-hmm.

Vit Lyoshin (09:20.133)
Hmm, interesting. I guess it's a it's a good place to switch to explain what your company does and what is pro what is the problem that you identified and and how you solving it. So can you tell us about that?

Tammuz Dubnov (09:36.58)
Yeah. so briefly put myself, CTO and co-founder of Autonomy AI. we are a platform helping product teams turn intent into delivery inside production grade code bases. what that means is organizations come to us wherever they are in their AI transformation journey. If they're early on, then maybe they use us just to understand the and how to write tickets better in a more agentic way, to using us to create prototypes, to using us to open PRs.

Vit Lyoshin (10:01.243)
Mm-hmm.

Tammuz Dubnov (10:04.985)
To finally using us to like totally automate everything and have our platform automatically look at signals from their analytics and then think up features, implement them and show them and and help take care of everything, or even like monitor friction and and try to improve the UX, everything kind of autonomously. and our whole thesis is everything needs to start from the code. The code is the source of truth when you're talking about product. It

shows you how it actually works behind the scenes. And it's not some mode of abstraction like a PRD, that is the version, the PM's like idea or ideal of what it should be, but the code may be something else. It's not the Figma, which is what designers want it to look like, but again, the product actually looks like something slightly different. So the truth is always in the code. And that's why everything we do always connects to your code base, starts with our modeling of the code base, our insights that we extract from the code base.

Vit Lyoshin (10:52.315)
Mm-hmm.

Tammuz Dubnov (11:01.891)
And then we can support you in everything else rooted on that actual truth.

Vit Lyoshin (11:07.001)
Hm, interesting. And you mentioned no code component. How does no code fit into all of that?

Tammuz Dubnov (11:14.275)
So the users are meant to be product teams, not engineers. They have their own tools. And as we support the PDLC, the product development lifecycle, you need to be able to experience your product throughout the whole flow. And we never want to show you code. You will always see a preview. The preview can either be like a quick prototype to ideate with you. It can be an isolated component showing you the component and all its variants and all its states. It could be your live app where we show you the part of your app that we just changed or that you're working on.

Vit Lyoshin (11:17.371)
Mm-hmm.

Vit Lyoshin (11:34.907)
Mm-hmm.

Tammuz Dubnov (11:44.174)
So there's always that visual element, always that interactive element. You're always seeing your product and what it's being changed on top of it. Sometimes it it can also create like a brief. So you may see like a text document we're showing you with like diagrams of what's happening behind the scenes. but most people, the way they use us, it it jumps right into that visual interactive, your actual app or a subset of your app that we're working on. and you iterate with it, there's no need for you to see your code. It doesn't really help you.

Vit Lyoshin (12:00.475)
Mm-hmm.

Vit Lyoshin (12:14.649)
Yeah, and that preview is generated based on the real code in in the repo, right? As far as I that's why it's exactly the truth. Okay, interesting.

Tammuz Dubnov (12:20.377)
Yeah.

It's actually so we model your code and then we also model your infra. So behind the scenes, every single task gets its own sandbox, which means it's like a a remote computer computer, runs your entire environment, like your actual code. And then we render your app running on a computer in front of you. So it's exactly the same as your app in production, or frankly, it's the same setup that your developers have that allows them to run the apps locally on their machines. And here it's throwaway sandboxes. So every task gets a fresh one.

Vit Lyoshin (12:48.656)
Mm-hmm.

Tammuz Dubnov (12:53.235)
you can go crazy and and do whatever you want there.

Vit Lyoshin (12:56.165)
Wow, interesting. I guess I would go crazy if I have a chance to play with it. Yeah. can we go through an example of like what PM would go through an experience, right? When they have an idea and when they interact with your system and and they get the final PR that is like ready to merge. Can you walk us through this process, what the tool is actually doing and helping PM?

Tammuz Dubnov (13:23.307)
Absolutely. I'll jump to the hard scenario where you you don't have a ticket, you have like an idea. so if you have an idea, we have something called discover mode. the we also have auto, so the the agent automatically chooses the mode for you. but say you you give us a a general idea of where you're aiming. the agent then goes to the code base to make sure it understands how and what is built today.

Vit Lyoshin (13:27.833)
Mm-hmm.

Vit Lyoshin (13:36.294)
Mm-hmm.

Tammuz Dubnov (13:47.684)
it goes to your analytics. We support a thousand plus connectors. So often people connect us to, let's say, a mix panel. Then we can look at user behavior and then also have that shape our understanding and our plan. And then we give you a brief. So the brief shows this was your idea. This is what we took out of it. This is how we do it. This is what you're asking for in a feature level, basically a PRD. And in the top right corner, there's a button with four options.

You can tell us export to a ticket. You can tell us export to a presentation. Or you can tell us build a sketch or go ahead and actually build this. So often if the idea is kind of fleshed out, we'll see people go straight to build it. If it's not, then we see people go through sketch and then kind of iterate. If you do sketches, then often it it gives you like multiple variations, and you can go in this direction visually or this direction in product flow and people iterate with it.

Vit Lyoshin (14:28.943)
Mm-hmm.

Vit Lyoshin (14:40.101)
Mm.

Tammuz Dubnov (14:42.175)
But then again, as soon as they're happy, they say, okay, build it. And then it actually does the code change and it opens a pull request. And then depending on the nature of your idea, often people have it create a presentation and then they share it with the team. And so that they can get alignment and people to know what's coming and know why it's coming. And that that thread of in task shows your your thinking process on why you need this. And so the presentation, the PR is here's what's been done. But PMs, it's really important to communicate the why.

So the presentation lets them communicate the Y. And I guess maybe the last thing that some tasks would need is you may send it off to like a designer counterpart. And then they can open it and they can interact with it as well and do like the design polishes. So we see RPM starts tasks, get to a place where functionality-wise they're happy. And then they call our designer, and then the designer keeps going and does all the visual polishes that he has the eye for. And then we send off that PR.

Vit Lyoshin (15:38.395)
Mm-hmm.

So people can actually also collaborate on the same task with different skills and everything. Mm-hmm. Yeah. Okay, cool. It's like a basically branch of the code and multiple teams, like almost like per programming. They work on the same thing until it's perfect. And then they send it out as a PR. Cool. Interesting. Okay. Yeah. And then also

Tammuz Dubnov (15:45.047)
A hundred percent. Yeah. It even sends you stuff in the Slack. So super collaborative.

Tammuz Dubnov (16:02.435)
Yeah. Exactly that.

Vit Lyoshin (16:08.535)
you mentioned that you have agent going through analytics and some other like customer interviews or whatever to figure out, but how does it figure out what is just people like theoretically asking about or thinking about, maybe this idea, maybe that idea, versus there is a signal that many people want that. How does it figure out what exactly like to help PMs to build?

Tammuz Dubnov (16:35.427)
That's a great question. so I'll talk about what people are doing today or w what we've seen pretty often. we see people opening like chat GPT threads and and talking about feature ideas. And then each thread is is kind of its own. And you can say that chat GPT like saves some sort of memory, but we see that that's really, really low fidelity, it's low quality. If you start a new thread about it with the same feature idea, you're you're almost starting from scratch.

Vit Lyoshin (16:39.472)
Mm-hmm.

Vit Lyoshin (16:47.323)
Mm-hmm.

Vit Lyoshin (17:01.051)
Mm-hmm.

Tammuz Dubnov (17:01.327)
And that's the gap, especially for us, because we're trying to serve not individuals, we're trying to serve an organization or an entire product team with I don't know 30 PMs or a sales team of like 70 people, and you want to get the insights from all of them. So the way we went about it, we call it spaces. Spaces is not your code, it lives above it. It's a knowledge hub that the agent can read and write from. So every time it does work for you and it creates a brief, it does some stuff and it writes it to its spaces. So there.

Vit Lyoshin (17:09.894)
Mm-hmm.

Vit Lyoshin (17:29.744)
Mm-hmm.

Tammuz Dubnov (17:31.13)
Know you have your a PM, you're going in with a feature idea. You will see in the spaces, not even you, the agent picks it up on its own. It has like a maps of content and the way it is built up. it will see that this same thing has been brought up by other PMs in the past, or it will see that there's been, I don't know, seven call recorders where clients have brought it up. and often it will just create its own ledger of ideas and then it'll start like keeping count and then like pulling the top one.

Vit Lyoshin (17:42.597)
Mm-hmm.

Tammuz Dubnov (18:00.64)
but then without you needing to to do any explicit work or even really thinking about it, the agent starts like creating their own knowledge graph of hey, this feature's been requested. Hey, this feature's coming up, hey, this is being brought up, I don't know, six times now. and that's how it's not like a one off session, but it's accumulative knowledge over time.

Vit Lyoshin (18:18.615)
Mm-hmm. So it's almost like a proactive user interviewing basically. You have an idea, you heard it somewhere, maybe, you start chatting with the tool, and it tells you also like, hey, ten more people asking about this in slight different variations. So pull it all together and build one that will satisfy everybody, basically. That's amazing. That that's pretty good.

Tammuz Dubnov (18:40.323)
Yeah, we've also seen it's like it's really cool actually. and we've seen times where people like start it and then they go away. Like I don't know, you had a feature idea and then you got busy and you forgot about it. And then I don't know, four weeks later you come up with a new a new task and it will surface that old feature idea and say, Maybe maybe it's time, like here's great insights you had there. Let's bring them to life. so it it's really cool. Actually, it it's pretty amazing.

Vit Lyoshin (18:43.71)
Vit Lyoshin (18:49.723)
Yeah.

Vit Lyoshin (18:58.437)
Mm-hmm.

Vit Lyoshin (19:06.265)
Yeah. Well f for you you sound it it sounds like for you it's normal you get used to it. But when I'm hearing this first time, it's pretty crazy and amazing because I think a lot of PMs actually need this information because they when they validate their idea, they spend a lot of time and a lot of investment going into just validating these ideas. But if you have a system that kind of on the background double checks on

On all your customer calls, sales pitches, or whatever, demos and everything. You already have a lot of this information sitting somewhere. And I think AI, it's a great use case for AI, right? To go through all of that and help PMs identify and surface and say this is a great idea. This is a supporting things from all these different people over the last six months. Go ahead and build it to satisfy everybody. I think that's a pretty amazing. Yeah.

Tammuz Dubnov (20:00.77)
Yeah, I I can even share like a cute story we have on our end. we have the system monitoring our our user analytics and coming up with feature ideas. It can like I don't know, whatever it wants to to do, whatever insight it got from that session. and we had one PR it opened that was just like adding like a keyboard shortcut 'cause it saw that our second highest event was people changing the layout and this and that way. and it went, did the work and opened the PR.

Vit Lyoshin (20:04.731)
Sure.

Mm-hmm.

Tammuz Dubnov (20:27.209)
and then our PM came all excited because she said, This is something I've been meaning to ask you guys to do for weeks and weeks, and we've just been busy. And this is exactly what I've I've been feeling within our product. But the platform came up with it on its own by being basically doing analytics on our events on user usage. and it's pretty amazing that there's within a product, like the next thing that's missing, it can be discovered in so many different angles.

and it's amazing that that she came with it from her angle and the the platform came with it from analytics angle and the events measurement. but it's really cool when the when the puzzle fits together, it just it fits together. It just makes sense.

Vit Lyoshin (21:09.175)
Mm-hmm. Yeah, yeah, that's great. I'm also curious if you have any sort of safeguards and we can talk about that. Like when PM is like building this feature, this new idea they want to implement, it do the do you trust automatically all these PMs? They can be like different skill level, different understanding of the architecture of the system. How do you do this safely for the whole product so they don't break or damage anything, you know?

Or or have a security vulnerabilities and stuff like that.

Tammuz Dubnov (21:42.868)
that is a great question. So let's talk about who has the authority to decide on what. the way I like to explain it is there's the why something needs to be done, there's the what needs to be done, and then there's a how, how it happens behind the scenes. generally you can kind of think of it as like the application layer is the what, it's what people interact with, and then there's all the the back-end business logic, and that's the the how the magic happens that enables that application layer. everything that is a how.

Vit Lyoshin (21:47.961)
Mm-hmm.

Vit Lyoshin (21:54.927)
Uh-huh.

Tammuz Dubnov (22:12.429)
That that needs to live with your developers. They you need their judgment calls. It's often has to do with complicated like schemas and and db decisions and architecture and a bunch of stuff you don't want your PMs to think about. And but the PMs do have the authority to decide on the the what, what I want this to look like, what I want the user to do here and there. and the beauty is that they can do that. we in our platform we let them do it totally, end to end. but the agent, since the agent knows the code.

Vit Lyoshin (22:26.48)
Mm-hmm.

Vit Lyoshin (22:40.166)
Mm-hmm.

Tammuz Dubnov (22:42.635)
It knows where that fine line is between I'm going to change the application layer to I'm going to be starting to touch back end. And this is sensitive stuff. And you want an engineer to make those judgment calls and not the PM working through whatever feature it is that they care about at the moment. So the safeguards are automatic. As you do your work, the agent will do what you want end-to-end to communicate your intent into a working prototype. But the PR may be only in the application layer and it may mock some of the back end.

So that we don't overstep. And then when we open the PR, there's a very clear description in the PR saying, Hey, dear engineer, I've worked out everything the PM wanted. This stuff needs to be figured out on the back end. I don't want to do it for you. This is where you need to go. But here's mock data so you know exactly how it needs to fit in with what's built. and then you you as a PM don't need to think about it. You again, you communicate what you want, you get it to a place where you're happy with what you have. And then the engineers do the back end if needed.

Vit Lyoshin (23:26.168)
Mm.

Vit Lyoshin (23:31.139)
Mm-hmm.

Tammuz Dubnov (23:41.815)
Or oftentimes what we see is, I don't know, eight out of ten times the changes you want will not require changes there. So the PR is just good to merge as is.

Vit Lyoshin (23:52.42)
Okay, so it's smart enough to pr to kind of also know what the backend looks like and make suggestions for engineers. But engineers always there to double check and make the I guess technical approval, if you will, of whether it should be merged or not. Mm-hmm.

Tammuz Dubnov (24:09.431)
Yeah, so th they push what gets merged and they can give comments and the agent will fix it. But if it's back end and there's architectural implications to the implementation, then it will leave it to the engineers to do. That's the stuff that the engineers wanna be doing. They don't want you to to do it without them.

Vit Lyoshin (24:14.353)
Mm-hmm.

Vit Lyoshin (24:26.427)
Yeah, that's actually pretty common, I guess. Engineers always like a little bit grumpy if you ask them to like do a lot of bug fixes or a lot of maintenance or a lot of like silly stuff. that's always been a friction. So and I I also know that developers, engineers usually like to do like rewrite. We need to rewrite this whole thing because it's all got so messed up we need to rewrite. So I think it gives them a little bit of a yeah, break.

Tammuz Dubnov (24:52.003)
They they Yeah. No, no, they love to refactor. We've seen a part of the the impact from from using our platform is typically engineers get to refactoring work more often. Because now that the PMs and the designers, anything they don't like, they can just fix. Like who finds the bugs? It's QA, it's PMs, people using the product and saying, No, no, this is not what it's supposed to happen, let's fix it. and now that they can just fix it. They they go to us, they fix it, they open a PR and they're on their way.

Vit Lyoshin (25:05.307)
Uh-huh.

Vit Lyoshin (25:09.051)
Mm-hmm.

Vit Lyoshin (25:15.92)
Mm-hmm.

Tammuz Dubnov (25:21.435)
and engineers get their bandwidth back. They get their time back. They can do that big refactoring effort they've wanted to do because all the the busy work of day to day is covered now.

Vit Lyoshin (25:32.28)
Mm-hmm. Do you find that this becomes more disciplined and and that PM now is taking accountability pretty much completely for this task and for this feature that they implement? Do you find people are more what's the right word? Like are more respon take more responsibility for this and they're more careful with this, or we need to have some more

policies and guidance and more be more careful of like adding those new features to the system or not not from technical perspective but from the like overall product management software development standpoint because now everybody can start making so much so many changes to this to the thing what what's your take on it how do you approach this or or coach the teams with this

Tammuz Dubnov (26:27.395)
so we've seen that the more people use it, the the more the appetite increases. in the sense of any PM you talk to probably has like forty or fifty things in the product that drive them crazy. And frankly those things are in the backlog and nobody manages to get to them. and now they can just get to them. So anything that bugs them that they just fix right away.

Vit Lyoshin (26:33.36)
Mm-hmm.

Vit Lyoshin (26:40.593)
Mm-hmm.

Tammuz Dubnov (26:50.999)
And all the little gaps in the user flow that they've always known about and always fought to get them in the sprint and lost that fight. they just take care of. so there's so much of that that that there doesn't need to be negotiation. People just get to the work right away. And then the in terms of discipline, what we've seen is people do like to share prototypes more than before. Like they have ideas. with us you can you can implement to a PR. You can also

Vit Lyoshin (27:00.326)
Mm-hmm.

Tammuz Dubnov (27:20.931)
Build a really quick like we call it HTML mock prototype. We call it sketching it, like you could sketch it out. and then because it's a collaborative effort, it's really easy to just share it with somebody else on the team and iterate with them. so we've seen that it's it's faster now to get buy-in from more people and align more people quickly. That's why we also added the presentation mode. because we've seen people like to like quickly present their concept.

Vit Lyoshin (27:32.111)
Mm-hmm.

Vit Lyoshin (27:44.165)
Mm-hmm.

Tammuz Dubnov (27:49.368)
So I don't know if I'd say they're being more more careful and like measuring stuff that the system does it for them along the way, but they are being more collaborative and they spend more time aligning and showing with others.

Vit Lyoshin (27:56.932)
Okay.

Yeah. Yeah. Interesting thing, I I'm sure you heard of this, but this is like Amazon's FAQ, PR FAQ idea that you start with the end, right? You write your press release and describe how you want this thing to work for customers, and then they basically meet between each other, collaborate on this thing to decide how they're gonna actually go now and implement this. What what you told

What you just said is basically you doing all of that press release and building the prototype to collaborate with your team and agree whether it's it's the right thing to do or not. and I think yeah, that that's that's a very good way to people see, interact with it and actually understand it better and faster what what what it's gonna look like at the end. Yeah, I think that's amazing. That's great.

Tammuz Dubnov (28:57.229)
Yeah. we also see that a lot of people don't always push it to the pull request. A lot of stuff get gets stuck at the idea phase, which is not bad. Like it's a part of going quickly and figuring stuff out. And then that gets left in the spaces that I mentioned earlier. So I don't know, four months later, if it's right, it'll naturally surface again. but yeah, people just they go through ideas and they iterate really, really quick. Not everything makes it to something you want to put in the product, but you suddenly can can get through a ton of ideas and mature really quickly.

Vit Lyoshin (29:02.106)
Mm-hmm.

Vit Lyoshin (29:05.968)
Sure.

Vit Lyoshin (29:15.398)
Mm-hmm.

Vit Lyoshin (29:20.699)
Yeah.

Vit Lyoshin (29:26.081)
Mm-hmm. You also have I forgot how it's called, but you mentioned before like an autonomous agent builder or or whatever it's called, you can say official name, which is basically doing all this research and half PM's jobs basically on itself and recommending what needs to be done and goes builds it. Can you explain how does it work? And and actually can you say the exact name, how you call that feature?

Tammuz Dubnov (29:55.3)
Yeah. So we call it our autonomous product delivery. and it it does basically what you said. It goes we're already connected to everything. We know your code. We've done a lot of work on it. So we know how to work within your code. And we're connected to a a bunch of data sources, whether it's like call recordings to analytics, to I don't know, Notion, your ticketing system. We can see your bunch of tickets and backlogs. We know just that frankly, we're connected to a bunch of stuff.

Vit Lyoshin (29:58.652)
Okay.

Vit Lyoshin (30:05.071)
Mm-hmm. Mm-hmm.

Tammuz Dubnov (30:22.831)
And then the question is, do we need to wait for a PM or designer to come to us with a task idea, a feature they want to do or something they want to change? Or can the agent leverage all the signals it has to come up with idea on its own and and then just get approval after the after the fact as kind of a proposition? the way I think of it internally is kind of delegating, like I'm delegating ownership of an area to to the agent. So

Vit Lyoshin (30:22.885)
Mm-hmm.

Tammuz Dubnov (30:51.215)
And it's also stuff that people don't really get to. Like that, that's the beauty. There's kind of the core and the peripheral in your product. And the core is what your CEO wants to talk to you roadmap-wise, but the peripheral is everything you need in, so a user can use this smoothly end-to-end. And suddenly you can like offload the peripheral to your agents. So we have one running on like let's let's say maintaining our analytics. So it goes, it runs every other day, it looks for any gaps in the analytics.

Vit Lyoshin (30:56.517)
Mm-hmm.

Tammuz Dubnov (31:21.089)
in the code, meaning is there some sort of user action where the code doesn't write and I don't know, send it off to a mixed panel? And if so, I'm gonna patch it so that everything's good. We have another one running on our accessibility and maintaining our accessibility. So we don't need to think about it anymore, not engineering and not PMs. That's already taken care of. We have another one looking for friction in our analytics and where users kind of hiccup or where user needs to do, I don't know, 20 clicks instead of seven clicks and automatically.

Vit Lyoshin (31:28.688)
Mm-hmm.

Tammuz Dubnov (31:50.692)
Patching that. We have another one that runs and looks for bugs and patches that. And suddenly the user's job, the PM, is to be the human on the loop. Hey, here's a a brief with pictures before and after what was changed and why I wanted to change it. And they say, Yeah, I like it. Send it to an engineer. And no longer being the the one to ideate on those small peripheral elements. Not because it's not the PM's role. Like you could say that's what they used to do.

Vit Lyoshin (32:13.393)
Mm-hmm.

Tammuz Dubnov (32:19.267)
But now they can focus on the core items, the stuff that the CEO cares to talk about, and not the bunch of peripheral stuff that the agent can just cover for them.

Vit Lyoshin (32:28.237)
Mm-hmm. Are you setting up those parameters yourself somehow? You basically giving some areas to focus to this autonomous product delivery agent, or does it like suggest and figures it out by itself?

Tammuz Dubnov (32:44.343)
so we generally recommend to kind of leave it as loose as possible and then the agent figures it out and builds its own places and the spaces where it creates a ledger and and tracks everything. we have a little one, we have a little template gallery of a bunch of use cases that we've found valuable and and our clients have found valuable. So you could take inspiration from that and go. And we also have a helper where it's an agent that talks to you and and sets it up for you and brainstorms and looks at your code with you to figure out what you want.

Vit Lyoshin (32:54.596)
Okay.

Vit Lyoshin (33:04.985)
Mm-hmm.

Tammuz Dubnov (33:13.539)
Basically you set the automation there and then and then it repeats however often you want and you just get to look at the results and say, I like it or I don't like it.

Vit Lyoshin (33:22.669)
Mm-hmm. What's your take on the decision making part of the job of a PM or engineer when we have agents like this now? what sort of decisions would you allow agents to do all the time and which ones you will never allow agents to do?

Tammuz Dubnov (33:40.335)
So I think decision making is basically the that's the role of a PM making decisions around the product. and everything a an agent proposes, it's the PM's decision to say yes or no. And if not, it can say why and try and iterate towards a yes. but there's nothing I would say you can't let the agent have a go at it. you still need to be the the gatekeeper.

Vit Lyoshin (33:50.331)
Mm-hmm. Mm-hmm.

Tammuz Dubnov (34:10.701)
So you as a PM have the authority to decide how you want your product to behave, what direction you want your product to go at. and while the agent can get more and more insights on on where you're trying to take the product, you for better or worse, you're the authority. You get the the credit or the blame. So you need to decide, hey, this is something I want in, or no. Oftentimes it gives you like nice ideas and you're like, this is like a soft no, let's get it to a yes. but it's on you. It's on you to know that your product.

Vit Lyoshin (34:17.093)
Mm-hmm.

Vit Lyoshin (34:28.891)
Mm-hmm.

Tammuz Dubnov (34:40.705)
is missing this thing that that users want and yeah, let's bring it in.

Vit Lyoshin (34:44.441)
Yeah, okay. Are there any specific skills now that PM should be paying more attention to? if they're gonna be like writing less PRDs and worrying about like specifics and acceptance criteria and stuff like that. What's your take on like skills? Are there any areas they should be learning more in in this new environment with agents?

Tammuz Dubnov (35:09.439)
so I don't know if you count this as a skill or not, but because everything goes so fast, we see that context switching is really important. The agents will have I don't know in in the span of what used to be one feature, you have six features that you have to keep track of and and approve. So you need to be sharp on the on the why for each one of them, but also be able to jump between them. You can still do deep work.

Vit Lyoshin (35:26.171)
Mm-hmm.

Tammuz Dubnov (35:37.338)
But the deep work happens so quickly that there's gonna be a lot of sessions of deep work. so I'd say that's something to be comfortable with. beyond it, in terms of skill set, it's also so so of course some PMs are like this already, but being as data-driven as possible, agents let you do stuff the right way because you can do work faster. So

Previously, people cut corners because there's that deadline and developers need to get it in and the sprint and whatever promised customers and so forth. Now you can go to sleep, you can wake up, you can basically forget everything you did till till now and say, if today's my first day is a PM and I'm gonna engineer the perfect way to build a product, like the absolute perfect way, and it's gonna have all the analytics, and I'm gonna do user testing, and I'm gonna show it to people before I take it in. And I'm you you know.

As a PM, you have your your gut instinct of how you want to work, but you don't work that way because there's there's constraints. And those constraints are going and and they're they're disappearing every day. So go back and and dream of what your ideal workflow is as a PM and don't be shy to start working in that way. It's it's doable now.

Vit Lyoshin (36:56.473)
Okay, so a lot of friction is going away with this agents. And you can really focus on the ideal process, how you envision this product to work, what users want, what you're hearing from them, and things like that, and all this busy work about writing requirements or going through and verifying, double checking everything, or whatever other busy stuff. Like I know user interviews take a lot of time. You have to go schedule them, re talk to people.

Then go through and analyze. It takes a lot of time that actually in some big organizations they have a separate group of people, user researchers, who just do that for PM and bring the ready-to-go analytics, right? And agents do the all of that now, and they actually do it pretty fast and quick. Interesting. Yeah, that's that's a new way. Like at the same time, you kind of have a lot of free time to do the things that you really need to do.

but that's at the same time you have to be careful now to make like those targeted decisions to focus on the most right things, on the exact right things. Interesting. And another thing I wanted to ask is about this like idea that everything is moving so fast now with these agents, they can make features like you said, instead of one, they can make six at the same time.

isn't it opening a door for like weak ideas or some nice to have's that many nice to have's gonna get through and not make an impact versus like the most important things are actually getting still prioritized and getting done? What's what's your take on this balance now that many like chunk basically is gonna show up in the in the product?

Tammuz Dubnov (38:43.615)
we focus on organizations, like larger organizations that typically have a roadmap and have backlog and know what they want to do. so that just naturally kind of organizes stuff. But the real answer is there's tension, right? There's tension between the person the builder running and creating work and opening PRs and the engineering team that's reviewing them. And every work that you do is work for them because they need to review and get it in.

Vit Lyoshin (38:51.941)
Mm-hmm.

Vit Lyoshin (39:07.91)
Mm-hmm.

Tammuz Dubnov (39:13.227)
and as soon as you get to this game, you will see that the engineering has to prioritize reviewing your pull requests. and it's not each other, it's not like they're close team, even within engineers, like between themselves, you see that they take a long time to review each other's PRs because there's work to be done and stuff takes time. so then if you open 30 PRs, you will see that they just don't get in. Like we have users that start off and we see that they like

I don't know, third day they open like 10 and then they wait three days and then they're why why isn't it live? Like, I I did this work, why is it not in? And then we slash their engineers say, Yeah, we have like a PRQ and it takes us time to get to stuff. and we see that naturally our users respect that and they know even if it took them, I don't know, an hour to create this amazing feature and it's out, it will take engineering typically a few days or sometimes weeks.

Vit Lyoshin (39:43.653)
Mm-hmm.

Tammuz Dubnov (40:09.261)
to get it through and in and and out the version. so they just prioritize because they know that there is this bottleneck that is engineering and they don't want to overload the bottleneck. so they prioritize what gets to them.

Vit Lyoshin (40:12.111)
Mm-hmm.

Vit Lyoshin (40:17.946)
Mm-hmm.

Vit Lyoshin (40:23.073)
Yeah, okay. So there is due diligence going on also before everything gets merged. So it kinda keeps it slow and keeps it double checking that like some you know lose and not what's the right word? Like some nice to have feature or something that is not right is gonna get into into it. So there is natural barrier there. Okay. I think it's a good one actually.

it's not bad because now I see more and more people talking about like, well, now everything is developing so fast, but how do you know which one actually needs to be developed? Right. So that's like a new problem right now for people to solve. Before they had so much to do, then don't have time to do it all. Now they have time to do it all. It just which one goes first? So it's a different problem.

Tammuz Dubnov (41:12.291)
Yeah. I I would say that the answer is typically in your analytics. Somewhere in there, there's like data that will help make the point or or break the point. and now that data is also accessible. So so it can help you validate ideas and and be data driven in your product decisions.

Vit Lyoshin (41:21.541)
Mm-hmm.

Vit Lyoshin (41:26.489)
Yeah.

Vit Lyoshin (41:31.417)
Yeah. So for the teams, like PMs or engineers who listening right now, and maybe they don't have this whole setup yet, but they want to, do you recommend them to start from or where do you recommend them to start from? Well, like which things to try to utilize agents, maybe even like copilot or something more simple than your system. But anyways, just to s get experience into in into this

New way of working.

Tammuz Dubnov (42:03.375)
of course I'd recommend our platform because it just makes everything kind of automatic and and default for you. but say you can't, the the place I would start, and it depends on the organization if if they allow you, is getting a hand getting your hands on the code. As a PM, or with most of the PMs I've seen, many of them don't have access to the code. And it's not important for you to look at the code, but as soon as it's there and the age and as you're thinking through a PRD.

Vit Lyoshin (42:07.141)
Mm-hmm.

Tammuz Dubnov (42:30.711)
You can tell the agent, hey, that double check what we have in the code already. And let's use that as a starting point and get that to where we want it. It can help you flesh out your PRDs. So the very, very first step we recommend anybody is let's see how smart your PRDs can get and start there. but smart PRDs are often anchored on your code because that's again the source of truth, and you want your your PRD to apply there. so see if you can get your hands on your code, even if it's read-only.

Vit Lyoshin (42:54.207)
Mm-hmm. Mm-hmm.

Tammuz Dubnov (43:00.611)
So that as you plan your PRD, then tell the agent, I know before or after, say, okay, here's the PRD, here's the code, let's see what's wrong with my PRD given how the product is actually built today. That's a great first step.

Vit Lyoshin (43:10.977)
Mm-hmm. Okay, great. And then also, let's say like five or ten years from now, given you building tools for the teams, where do you see like a perfect team or maybe not perfect but normal team ten years from now? How do they develop a new product? What what's your vision?

Tammuz Dubnov (43:32.88)
okay. It's gonna be I'm gonna phrase it weird and we'll figure that together. the product team is gonna be figuring out the the direction the product should go, no longer figuring out the nuances of the product because AI can figure out the nuances. I I called it peripheral earlier, but you can figure out the peripheral, make a product robust and everything, and you just need to figure out the direction.

Vit Lyoshin (43:39.723)
Ha ha ha.

Tammuz Dubnov (44:01.024)
And then a really successful team will figure out how they can insert, I'll call it data traps within the product that can hint to them what the next thing is that you want. Interviewing users on what they want next is problematic. We all know that. And but measuring their behavior and figuring out what they want is not. So if you do

Vit Lyoshin (44:16.785)
Mm-hmm.

Tammuz Dubnov (44:25.507)
a a product and and you put in some I'll I'll say Easter eggs of buttons that that are like coming soon after you press the button you like no no this is coming soon it's not out but from that you have user behavior knowing people will click on this people want this and have it throughout your product. So yes the core functionality is in and then there's like little data sources of stuff people can interact with that will tell you they actually also want this and they clicked on this which means they actually do want that. And you build a sort of engine that that is

data driven to tell you what the next thing they want without you having to ask them

Vit Lyoshin (44:56.471)
Mm-hmm. Mm-hmm. Okay. Interesting. It's yeah, okay. Autonomous system basically to help you interview and process all the feedback, data driven, recommendation, all of that, feeds back to the PM or decision maker, whatever the name will be ten years from now. And that person will make decisions to what to where to go next. Okay, cool. I'm looking forward to that. We'll see.

I guess everybody many people are have high hopes for the AI and how it can help us build better products and stuff like that. So that's great. All right. Yeah, okay. Thank you, Tamus, for the information. That's been great. And I think it's promising to see what you're building and like I almost wanna go ahead and try to build something myself now. but

Before you go, I have three more questions. This is my innovation QA for every guest I ask. Those are the standard questions I ask everybody. can you define innovation in a few words?

Tammuz Dubnov (46:09.599)
yeah. I'd say i innovation is solving a problem succinctly. it's being able to take like the the core gap and and actually address it without a lot of fluff in between to try and mask the real problem. I think that's the the most of the the innovation or today I think it's accumulation of a bunch of microinnovation. That's how people get far. is finding the biggest gap you have today, the biggest hole, and and solving just that.

Vit Lyoshin (46:32.209)
Mm-hmm.

Tammuz Dubnov (46:39.127)
Not stuff around it, just that.

Vit Lyoshin (46:39.569)
Mm-hmm.

Okay. And the second question is which one innovation do you think changed the world the most and why?

Tammuz Dubnov (46:50.381)
That that's a hard question given I'm in AI, because obviously I'm gonna say LLMs. that's an accumulation of just a ton of microinnovations that suddenly like had emergent behaviors and and brought us where it's bringing us. but I think that's gonna totally reshape the world and and the underlying of what's driving the world. so that that's the the innovation, the modern day innovation that that I'm all in about.

Vit Lyoshin (46:55.845)
Ha ha.

Vit Lyoshin (47:15.481)
Okay, sure. And the last question is which one thing like a software technology system tool that we use today we will be laughing at ten years from now?

Tammuz Dubnov (47:41.881)
That's hard one. I don't have a a beautiful answer. I think the place that's gonna be reimagined is all the points that we need to align today. I think a lot of different modes of communication are meant to help us align. like emails are meant to help us align, phone calls are meant to help us align, meetings are meant to help us align. I think in the future that will happen less. I think again coming from from AI and agents, I think agents will align for us and then we'll just

be updated on the final results. I can't put it on which technology itself will go, but the amount of effort and time that us humans working in a big organization with lots of counterparts, the amount of time we invest today in in in meetings and efforts of effectively alignment, I think that's gonna fade out. and we're gonna be shocked at how much time in retrospect we've burnt on just alignment all the time.

Vit Lyoshin (48:34.961)
Mm-hmm. Yeah, well looking forward. I hate meetings. So looking forward to that. Yeah. All right. Thank you very much for your time. It's been great talking to you.

Tammuz Dubnov (48:39.44)
Yeah, as do I.

Tammuz Dubnov (48:48.929)
Absolutely. It's a pleasure. Thank you for having me.

Vit Lyoshin (48:51.803)
Sure, thanks, bye bye.

Tammuz Dubnov (48:54.252)
Mm.