So thanks for having me at the stage Who of you has been using Cloud Code or Codex? Okay, great Who of you who have raised the hands are developers or are your product managers? Great! Lovely to be here. So my talk is about the Great Reset how AI transforms product management and UX and everything. And to get into it, who am I? Those of you who don't know me, I'm Björn. I'm the managing director of Mayflower. We are a small company, 75 people here in Germany. And I'm a retired developer who is now doing wipe coding. So I have a lot of fun coding until 1 a.m. 1 p.m throughout the night and my company we're building AI and data solutions for the enterprise so we're training models small models we're doing systems real-time voice systems deep agents I will come back later what that is and helping companies become AI organizations so giving you the capability and transforming yourself to leverage the full potential of AI, which is not so easy, to be honest. And my first question to you is, how does it feel to be the slowest in the room, besides the developers here? So throughout the talk, always think about this question. And yeah, who of you is using JIRA, Confluence, Linear, Notion. Okay, who's using Figma? Oh wow. Anyone organizing via SAFE, the scaled Agile framework? Okay, so the diagnosis is you're malformed and you need some kind of reset. And so let's compare to the others, compare to the laggards. so just joking compared to laggards compared to the experimentators and compared to the really AI innovative companies if you look at the US you see a lot of folks there who are doing incredible things and before we dive into it the basics what do you think is an AI agent just throw into just say yeah so there are no wrong answers obviously any other answers Okay, so my definition of an AI agent is a model which has rules, which has tools to your internal and external systems, you know the words MCP. The system, the agent has access to a system called bash, who knows what bash is. That's for the hackers, that's a common line things. because it turned out all those models are really incredibly good at invoking comments on your computer. You have skills. I was laughing at the Slido answers. Use skills instead of MCPs. That's great. So use skills instead of MCPs because it doesn't pollute the context window of the model. The problem is the models are so big. and they have such small context window with Claude I think around 200,000, 190,000. Sometimes with Gemini we had 1 million I think but the model itself was not so good for coding. The important thing is the dynamic task list. The model or the harness around it is incredibly good. at creating itself a task list and that's the essential tool for an agent itself because you have the problem that when you go into ChatGPT or Cloud itself or other toolings when you just throw a question in then you get an answer but that's only question answering but what about long reports long research tasks long things where the agent is running half an hour, 45 minutes, two hours doing a deep research in the internet. Without those guardrails, without those dynamic task lists, the model is drifting. So the answer is bad. So you need dynamic task list and models are very, very good at thinking and reasoning. So that means before a model gives an answer, It gives you the thinking and reasoning. Sometimes in Claude you can expand this in the UI and usually it's hidden away in the answer because normally the user tools you are using, they won't display the thinking to you. But it's inside the model and you can display it and you have the reasoning. So the model is thinking about what does the user want from me. And when the model gives an answer, it thinks about what is the answer. We use this in the agents to steer the system in different directions with the thinking and tell him, oh no, that's wrong, you have to self-correct. Models are very good at thinking and reasoning and the frameworks around that help the model to reposition itself so that it steers in the correct direction. And also what's good at, and if you want to remember two words out of the session, it's self-critique. So whenever you get an answer, say to the model, no, that's wrong. Please self-criticize yourself and come back with another answer. Maybe you've heard about the Ralph Loop, Ralph Wiggum. that was during Christmas real hype around that I think it will not be a hype anymore because Cloud Code will have dynamic task lists and swarm mode in one of the next versions so they are baking it into it and we have some kind of loops inside it so whenever you get an answer try to repeat those no that's wrong please think again, please research again, come back with another answer. Then the model comes back with another answer, then you say again, no that's wrong, please think again, come back. After four, five, six, seven repetitions you get a really good result out of the model. In late 2025, 2026, maybe I show it to you, we have sandboxed workspaces. the model itself is trained into executing comments on your machine but that's dangerous on my laptop doing that so you have to sandbox that and usually in some containers or virtual environments and stuff like that and who knows about that who has heard about sandbox whitespace because models are very very good not only at executing comments on your computer but also at writing code themselves and if you think about products you build for your users think about having the model write the code itself in the background to give you a short demonstration about this and then I will continue with my talk it's spontaneous that's a third tooling we've written and it has a sandbox underneath for the techies it's built with deno and vasm so it's secure by default and i type it in german so that the demo will work hopefully and maybe if we are a bit risky of course it should know english so um give me the ergebnisse of the sales of the last quarter as a diagram models can't do diagrams but they can't write programs and it's a bit slow but you see it's writing code. This isn't pre-configured it's just a problem and the model is thinking oh yeah I cannot give the answer but I can write code and so I write code and it's inventing some sales numbers because we haven't put the database in it and now it's giving me the results. and there's a diagram and then I say, make the presentation for the presentation as PDF with all the details. Create the presentation for the C-level with all details as a PDF and then it writes a program to create the PDF. Of course it invents some data and this all happens in the sandbox. So the real power for models and that's what you have to think about for your own products. The real power is not only doing a call to a Gemini, Claude, Entropic or OpenAI, the real idea is to put all these things together to have some really powerful systems and to create some really powerful features inside your application. so that you don't have to or your engineers don't have to write every line of code so that you have an autonomous system that's able within its guardrails to follow a way to follow a task so do this a long way calling a lot of tools calling a lot of MCP services and come back with the result that was done so continuing with the presentation so that's an agent and sandbox workspaces is really crucial for having large long-running agents when you build this now what has this to do with product management you may ask So there's a famous quote from Eng Drang who is one of the AI pioneers in the US and co-founder of Coursera. He said engineers are 10 times faster and product managers haven't sped up at the same rate, now they are the bottleneck. And that's why I asked you how does it feel to be the slowest in the room. And the question is not hire more PMs because the developers are so fast now, it's how do you speed up in your work. So spoiler, it's by working closer to the code and not farther from it. So you don't have to write software of course, don't get me wrong, but you have to incorporate all those principles of those AI based software engineering into your work as a product manager, as a product owner when you're thinking about the products. And we have fresh data. AI is over delivering. There was a survey from Lenny Ratschitsky of December 2025 with 1,750 respondents. And they said 55% said AI exceeded their expectations. More than 50% save at least half a day per week. Who is in this from you here? Saving at least half a day per week. Great. Saving more than one day or two days per week. So half-time working, that's great. Founders save 50%, nearly 50% save more than six hours, so one day a week. And Kevin Weil, who's the vice president at OpenAI, is saying, the AI model you're using today is the worst you will ever use because AI models get better and better every other month, every three months, every six months, they will get better. There was an inflection point, I think, last year around end of October, beginning of November, when Opus 4.5 from Entropic came out. And it was blowing everybody away because it was a really, really good... model for software development the codex is also good but opus is really better i would think so the paradox visualized engineering got faster by 10 times from 2019 to 25 and at the same time product management only got 1.2 faster in this time so the paradox is the parents are getting faster and software engineers are getting ultra fast but the whole system is appearing slower. It's not only organizational bottlenecks. We've heard about OKR, we've heard about Agile, organizational stuff and things like that and oh by the way AI is destroying Agile and Scrum. Who's talking about Agile anymore within AI? Nobody. Because as a developer I have my own team. Five agents, ten agents. 20 agents. I only do my scrum of scrums with my human teammates, not me because my colleagues don't let me work in the project anymore. But yeah, you get the point. And how do we work? There's a feature when working with code and when maintaining the code in so-called boxes, there's a feature that's called work trees. Whoever is using Git, it's a kind of a version control system for your code, for the code of your engineers. And there's a feature in Git where you can basically replicate the code version, the current version of your code several times. So that's the work trees. And as a developer, when I want to implement a feature, I just spin up three different work trees in the same directory. So they get their subdirectories on the computer. and I work with three coding agents in parallel on those worktrees. So on the same code you get the point. When you want to implement a feature you can say okay we have some feature ideas, three different versions, let's spin up the coding agents in three different worktrees, let them do the work. of course i'm oversimplifying let them do the work and let's look at the results and then we look at the results and maybe test it a b testing we can also do this automatically so we don't need any humans anymore and then the winner will be merged into the the source code so we can build three features or five features or ten features but three three to fours that what the human brain can work simultaneously by AI agents and while those agents are building what do the human engineers do? Of course they are steering the coding agents because AI is not magical I have to say AI makes mistakes the coding agents make mistakes but we have time now we're not just bumping on YouTube or playing Tetris or Candy Crush or things like that. No, we're just doing the next features or in a second project doing features and we do this all the day. So keep that in mind for your product work because you can adapt this to your work in my opinion. So what's the speed of AI native? OpenAI built the Zora Android app in just 18 days with two to three engineers. I updated the slides. You maybe have heard of Cloud Cowork which is basically only Cloud Cowork built into an Electron app, wired into the normal Cloud desktop app and that's why they built the Task Agent system because it is also in Cloud Cowork. Cloud Cowork, the first version was built in a week and put out to the users of course, hey that's a research demo, okay, fair enough. But I organized, I think, thousands of files in my downloads folder on my MacBook with Claude Kovac because it had a connection to my Chrome browser, it had a connection to my file system, and I just typed what it has to do. And it did it and it made mistakes, but it learned from those mistakes because it could see what it is doing in the Chrome browser and could self-correct. That's the self-correcting things of agents I was speaking before. And so the question you have to ask is your organization built for this potential of speed? How are features rippling down from the idea cave to the product management? Maybe you have to talk to your stakeholders. Who has to talk to stakeholders for getting budget investment stuff? Yeah, that's slow because sometimes you have to talk to five stakeholders, ten stakeholders. If you are in really, really big companies, you have to build wireframes. Who's building wireframes? Maybe it's time for an upgrade because wireframes slows you down. So if you think about Kanban, the way of optimizing the flow in a system is from idea to production and to optimize every point. And with generative AI and especially AI and coding agents. It's the first time we are able to do this at lightning speed. So the problem is the absorption problem. So 75% engineers, that was a report from 2025 I think, use AI assistance companies, but companies see no measurable delivery improvement. The PR review time, so that means whenever when an engineer does a change into the code base he doesn't do it on mainline except you on Google. That doesn't work anymore in the AI world when you produce code at lightning speed. Who rarely use that anymore? GitHub, that's the number one code base in the world, is full of established maintainers of large open source software projects who say, no, we don't want AI generated pull requests because there's so much slop in it. And that's a fundamental problem because the architecture, how we developers work, is not true anymore. So we have to change too. So the system eats the speed and there's a visible trend in the US. It's called unsupervised software development. No human is looking at the code. And I think eventually we will get into this in some years here in Germany, smaller companies. They can do it easier than larger companies, large enterprises, where you have those organizational structures and stuff like that. But the trend is clear in the US, you have a lot of unsupervised software development. And the result is not, oh, okay, now we need only product managers who write their ideas down and then there's the automated code factory. who produces code and that's working on that's not true but you get the point of the trend and that's the same also for product management and you can question yourself where are you I called it the 70-25 reality so 70 to 75 percent of the current landscape is traditional landscape so like JIRA quarterly roadmaps you can exchange Jira with Trello, Linear, whatever tool, it doesn't matter. Quarterly roadmaps, approval chains from top to bottom, executive sponsors here, executive sponsors there, manager here, manager there and that's probably most of you and that's still okay. That's fine folks. And 20 to 25 percent are hybrid. They use ChatGPT for research but all processes remain. We're experimenting but not transforming. It's like putting your toe into the big lake of generative AI. And then there's four to five percent who are AI native. So they have continuous everything. Prototype first, delivery first. And these are the pioneers who are two to three years ahead. And what AI Native looks like, of course, yeah, we could self-criticize this because Lovable builds a product for this world, but they got to 200 million ARR in under one year with just 100 employees. And Elena Werner, head of growth at Lovable said, we have to refine product market fit every three months. So the cycle is every three months. The landscape is shifting that fast. And from my personal point of view, if you use something like Lovable, and you want this GUI type of feel, just have a look at Replit because I think Replit is more in the neighborhood of software engineering than Lavable is and it's easier to deploy these services. So 60 to 70 percent of traditional growth tactics no longer apply into the AI world and those companies operate on a completely different clock speed. So it's not ultra-lightning fast because life is normal, but you get the point, you get the idea. And we have a real gap here in Germany, Austria and Swiss. And I think we may look from a negative point of view on that, but we could also look from a positive point of view of that. So in the US and UK, you have AI native workflows who are already established in Germany, Austria, Swiss. We are years behind in adaption. and i told it on on the conference digitale leute summit at product con in san francisco 2024 so two years ago we had ai native product teams while in cologne the pm1 summit had the session jump start your ai journey so you see the difference that that's really large so you're still starting while the us is running and this gap isn't just a problem it's your competitive advantage if you move now if you change something. And that's also a case study also from Lenny, I think from December. Zevi Arnovitz, Product Manager at Meta, he has no technical background and his engineering teams asked him to teach him how he builds with AI. So he uses Cloud for planning, Gemini for UI, AI models, peer review each other's code. They built a study made app live during the podcast. He said this might be the best time to be a junior in tech. I haven't been developing software for 15 years and I have the fun of my life. I think I have more advantage than a senior engineer because my brain is a blank canvas and I can experiment with that. While the senior and the architect says, oh no, they are doing errors and you cannot do that and cannot do this. And I say, yeah, that's fine. Then I prompt, okay, fix this, fix that. And oh no, look at the architecture, do this, do that. I talk to my agent and you should talk too. So what does prototype first mean? That's why I ask who's using Cloud Code or Codex. Use Cloud Code or Codex or other coding agents as a product manager. for UI prototypes directly in the Git repository so directly in the code don't fear that pair with one of your engineers tell them well give me some small space in the code space some small directory maybe in a dot directory which is invisible everybody who is technical knows what I mean and during conversations with your stakeholder you could use the coding agents to generate prototypes on the fly, to generate variations of the UI, to iterate live with feedback during your talking with your stakeholders. And ideas become credible when they exist as software. You know the believability curve, paper prototypes are nuts, they are nothing. But working software is everything. because we can imagine, we can touch it, we can feel it and if you get to the touch, to the feel in seconds instead of hours or days because you have to wait for your engineering team you're doing a better job as a product manager. So the idea is to get from idea to prototype in hours instead of weeks or some days instead of weeks. Working software is always better than PowerPoint decks or Figma, Miro, whatever you call those tools. I said AI natives, they are doing the feedback loop. I have an Agile background, we're doing Agile since 2005, so 20 year anniversary. The basic idea was always in Agile to tighten the feedback loops between discovery and delivery. And when I give courses to companies, they always ask how many percent discovery and how many percent delivery should we do as a team? This doesn't hold true anymore with Generative AI because you can merge those two together. That's why I told you work as close as possible with your engineering teams together, pair together, write code through software tools. even if it's lovable or other things. And AI-native teams, they are circular, they are continuous. They have discovery, validation and delivery all connected together in one tight feedback loop. AI generates automated tests. AI deploys automatically. AI measures automatically. AI reports automatically. AI draws conclusions. and you can also do with browser control, I'm sure you have heard that browser use you can create virtual personas and you can say test this new feature as Lisa from the accounting department and give me your conclusions when you test that and then the AI agent spawns a Chrome browser, headless Chrome browser clicks through the new features as Lisa from the accounting department or the AI thinks as Lisa from the accounting department and gives you the result and you are more like an orchestrator drawing yourself your human conclusions and merging that with the conclusions from the AI which is sometimes good and sometimes bad and everything runs in parallel and everything informs each other and then you get speed of light That's an example shameless plug I just wrote a software for fun because I'm an avid Confluence and Jira user who's using Confluence and Jira Cloud I hate it, I really hate it Who's hating it? But I wanted to write a tool just for fun We were on holidays on a cruise ship I took out my mobile phone, connected to my home lab at home and spawned cloud code and said, well, what Atlassian command line interfaces are out there? Because the rationale is this. As I told you, the AI models are pretty good in executing comments. And how would it be to have a tool who does a bi-directional sync of confluence pages? to my local hard disk, converts the Confluence pages to Markdown and if I change the Markdown file it will get uploaded back to Confluence including macro support and things like that. And then I went crazy with Claude because I didn't write any software in the first run. I just said well do a deep research. Spawn several research agents, sub-agents. The research objectives were research features of other Qlik Commons, which pain points do they solve? Do the same for Atlassian marketplace listings. Go to the Atlassian marketplace and crawl it. Do also go to Reddit. What are the customer complaints about? Confluence working in the wiki system, converting that, what are the complaints of Confluence administrators when they have to do the heavy lifting between pages or the Jira people. What about PDF export? Why do I have to pay for very expensive large commercial products? I think you know who I mean. Then I told okay scan that, prioritize, extract the jobs to be done, extract the pain points and give me the feature ideas my tool should have. Also research because although it's easy to produce software I want to have a meaning so is there any other product that has this bidirectional sync I told you because it's a core feature of the tool and then I said into the prompt always stay factual always stay current as of December 2025 because sometimes the models maybe you have experience that they search in for 2024 or come back with old results who don't test the truth anymore or stand the truth anymore always provide urls of your findings because i want to manually look at it do a self-critic before presenting that revise your research result and then present to me and write it down in a markdown file because i have to manually review it You can go to the URL into the specs directory. There's a bunch of research stuff to other aspects for the tool. That's the way I want you to leverage with your agents. Nevertheless, if it's a coding agent, if you're asking ChatGPT or if you're asking Cloud or other tools, do this research loop. Say, no, that's wrong. Go back. come later, self-criticize yourself, give me the URLs like perplexity does for example and then manually come through the document and then do a next round and do always your repetition. So the RALF loop for research and for product management is also true here. And your bad future for 2027 will be that nothing happened. So approval flows will be 10 hops or more. AI is barely adopted in your company. So it's like we rolled out Copilot. Yeah, Microsoft Copilot. We are AI. Product manager remains the bottleneck. Developer teams are frustrated because they have the tools, they know how to speed up and the system around them is very slow. And then there will be a brain drain, so the best will be for AI native companies. And that's the thing you don't want to have. So how to keep up? Some practical advice. First of all, like always, learn the principles. It's not about the tools sometimes, but it's not about the tools, it's about the principles. Carve out time to experiment for yourself because you need time. We all have a private life, family, friends, going out, etc. But in your work time, you have to carve out time for yourself to do those experimentations. Brainstorm with your engineering team. Sit together, pair with a developer. Go into the coding mode to learn this, to learn how they think, to learn how the thinking changes when using coding agents. Do discovery and delivery and brainstorm. How could we... merge those two processes together to get the full potential. So do it, experiment and especially in parallel. And of course the old Agile at Age inspect and adapt. So learning the principles. First of all, Markdown is a new notion. So throw away your tools. It's learning Markdown. And I don't care if you're using Obsidian or whatever tool you use. It's just a file on the file system in a very easy readable file format that's easily to feed into the agent. Bash is all you need. So open up your terminal on your computer, learn not how to click but how to type. And then as we had in the Slido, skills versus MCP versus sub-agents. So MCP has the problem. or had the problem until the end of last year, until December, that it bloats the context window of the agents. And the number one problem of models is that when you're near the context window, when the context window is full, like 80, 90%, the models are drifting. The results are not good anymore. They tend to hallucinate and stuff like that. And MCP, fills up the context of the model but skills do not skills who knows who skills are? Okay skills is basically a zip file with an markdown with instructions and software programs you can add like small python scripts or small type scripts that the model can execute and basically in a skill file you say you have a headline a header where you say this is the skill for as you have seen for writing creating PDFs or creating diagrams and then you have the complete instruction how to create diagrams and the thing is that when the model itself or when the agent uses the skill or incorporates the skill it only incorporates the headline so what is the skill named what can I use the skill for and that's all and this counts only several hundred tokens and so the agent knows when a user wants to create a diagram I call the skill and then it unfolds the skill and then it fills the context with the instructions of the skill so it's basically a very clever way to package instructions for the model Then you have the so-called agents.md or claud.md that's basically also a text file with instructions for your model and whenever you make a request to the model the content of this file will be feed into the request so it's kind of a prerequisite document. Hey you know we're in the ATL-CLI project and it's for for doing this and that with Jira and Confluence and hey my testing philosophy is not only doing unit tests, no we don't do mocks, we don't lie to ourselves, of course we do end-to-end tests every time I ask you, no don't commit to git on yourself, please ask me before when you do that, yeah and so on and so on and so on. So basically you have a junior a junior developer and you have to give him instructions because if not he tends to forget that and that's basically the rules for doing that. Then you have the prompting, the guardrails I told you like when you do research for example, okay if you do the research please self-criticize before you present that to me, please be so kind, do that two or three times and then you come back to me, okay? So that's the prompting. You have some magic words, half a year ago we had those magic words, think hard, think ultra hard, ULTRA HARD! Hardest, hard, hard, hard. And then Entropic said, yeah we heard you, you don't have to scream at us and you don't have to say ultra hard or think hard because the model has in its system prompts, you have to think hard. So they fixed that. Then you have the RALF looping which basically is the loop so that when a model does a task you say yeah that's okay but go back look at the code changes you made did you do all the tests have you looked at Claude.md do a code review again and just fix that do me a favor please So if you get one tip out of here of this task, it's repeat yourself. Do this the next time. So we had this before. Whatever you build, your product or whatever you are using, a coding agent, have this in mind. A coding agent who is able to do long-term research, long-durational tasks, has to have these things. in order to give reliable results. So build these into a coherent system and learn from the coding agents. So the best software systems mimic the coding harnesses. So if you have ever heard of Langchain, Langgraph, DeepAgent, TypeScript, these are one of the best open source tools out there and they mimic the behavior of the coding agents. What else? Agents should have a memory. We divide between conversational memory, episodic memory and long-term memory. So basically that means conversational memory is the agent remembers what you said three turns before and long-term memory is what are your guiding principles and things like that. I told about agents.md, claude.md and that's some kind of long-term memory because it will always be fed into the context and whenever you build a system of course you have the problem memory grows and grows and grows and you don't know what parts of the memory have to go into the prompt the user is giving to the model so that the model itself is not hallucinating because the problem is if you give The prompt to the model and your large memory bank of gigabytes of memories from different contexts, from different domains, it's like a human. It says, what's happening? I don't know. And then the results will be not so good. Give skills to the model for UI components. So whatever you are using in your product, so for example, Shotzi and UI, React components or stuff like that. give the models the context and that's why I said that as a product manager you can adapt this so give your agent you are using your tool you are using for your product management work your AI tool give it as much context as possible because then the results you get will be better you can give skills for drawing so there's a tool called pencil.dev I think it's commercial now it was it was around Christmas or shortly after Christmas, it gave the coding agents a drawing board. So basically connecting to a server and having a skill which explains, okay, when you want to draw you have to do this to do this, give me this comment, give me this comment and then you have a drawing board for the agent. Because sometimes we as humans, we tend to talk in a visual way. but the models are talking in text textual way and in product management as I told you the example before you test these new features as Lisa from the accounting department you use browser automation or computer automation to test your software and to tighten the feedback loop to the agent so you don't need any manual tests anymore you just automate this and of course you have to do backups because when something goes wrong it's great to have a backup. Rules, rules, rules. Agile people don't like this but basically LLMs work very great when they get a lot of rules, really a lot of rules. There was a question on Twitter to Boris Czerny who is the creator of ClotCode I think it was two days ago or so and somebody asked well I would like to see what's in your Claude.md how big is it how large is it and he said well it's 5.3k tokens and I don't know what 5.3k tokens are how many words basically 0.75 words and I said I got into Claude and said well your creator said his Claude.md is 5.3k words along created an example Claude.md which has the same size and then I got such a book of Claude.md and it was to compare if my Claude.md is going into the right direction do I have enough rules because sometimes we think oh no not the 100th rule the 200th rule that's not good and yeah it's good so give a lot of rules And if something bad happens, tell the agent, oh, don't do this anymore next time. So add this to the claw.md. I have another software fun project and I built a learning loop inside claw.md. I had the rule inside it and said, whenever you finish a new feature which was tested correctly with unit tests, integration tests, end-to-end live testing, make a summarization of your learnings along the way and summarize it in such a way that you add it to claw.md and with every turn I get a feedback loop into my claw.md and it grows and grows and grows and the software will be better and better and better and of course rules rules rules and the advanced usage is hooks hooks is some kind of stop mechanism for the agent system so that means when you can pretty configure the agent systems basic example when the agent wants to commit the code it thinks oh i'm finished i made everything now i can commit to the code and the hook says no you can't the tests are not good you cannot do this stop here so it's basically some some kind of security mechanism for your your results and of course do backups please carfai tom to experiment I said yeah maybe four hours a week or every other week maybe six hours eight hours depends on your organization what you can do of course install different tools so everybody is lucky who is in a small organization or in a big organization where he can install everything but oftentimes in big organizations you can't do that but you have to experiment because AI, generative AI is a field where every day new things come up, new optimizations, new ideas and of course you don't have to do this every day but maybe every week and learn out loudly. With the learn out loudly I mean learn out loudly internally in your organization, have group conversations, have internal meetups, go to external meetups, show the code. do life hacking together with your developers as a product manager to get a feel how those models are in those agent harnesses. Save your receipts, grow and hone them because tomorrow maybe some of them will be invalidated. And to give you a concrete example about a merging of a discovery and delivery loop, imagine this. There's a morning briefing. The agent knocks at 7 a.m. Throughout the night the agent has built insights and creates tickets for this. So it receives the user signals from your application. It receives the error reports from all those technical... systems your development team is using and in the past you were not interested because you said oh well that's development stuff I don't care about it I'm a product manager I don't want this code thing yeah but these are useful signals you can incorporate the agent has a briefing from his deep agents who were searching over the night in the internet about your competitors What did they do? What did they do in new features inside their systems? This system also scanned new customer wishes. So what did your customers do in the customer emails scanning that? So the system is clear. So basically every signal you can get you feed into the agent and he does a morning report and says okay these are the top 10 complaints and top 10 things and Before AI you use this and talk to your developers manually or you talk to your manager and say oh look how our system is doing great compared to the competitors but with AI the system itself you review and change those items of course manually you go through the lists and say okay no yeah that's fine but not now for the product, we do this later, but hey, these 10 features are great. And you mark them, you label them because the agent did not only a morning briefing, it also created the tickets in your JIRA, linear, whatever tool you have. And you pair with an engineer and say, okay, let's do those three. And you mark them as ready to implement. And before AI, you had to wait, the development team had a large backlog. Maybe in two sprints or two iterations we can do that, yeah. But now you say okay let's do and the coding agents are coming because the coding agents are scanning the tickets with the label ready to implement. In the ticket the agent has described two different variants of the feature idea and you course correct it of course manually. and then the coding agent spawns three different work trees of the system, develops the system itself, of course you have to have a lot of tests, etc. etc. and they develop those features and say okay ready to go you can test it under these URLs and then you grab a coffee with your dev teammate you look at the features and say okay from these three features we do number two and then it will be merged into the code base. And that's not a fantasy, that's true. This is already happening. And the next one is, of course, you could automate UX. So why should any human work in Figma anymore? I don't know. You? Collaborative. Yeah, of course, collaborative, maybe together with the machine. Yeah, design system, but the design system usually in an optimal way lives in the code. So you don't have to use those manual tools anymore. Of course you can do it but you can automate that. You can automate your UX, you can automate everything because the engineers they tend to have their design systems inside the code because no engineer wants to think about all this calendar component in the system. How did it look? So you have things like storybook and all those design systems. and they live next to the code of the application and what is in Git is automatable and so you have access to that. You can also automate the end-to-end UI testing as I outlined and then you have automated the full software development lifecycle. My impression is that in the US the companies are working on this automating the full SDLC from IDEA to delivery. So what are some helpful resources? Of course Lenny's newsletter, there's a bunch of interesting ideas there, great interview partners. Go to github.com, grab an account, search for awesome cloud skills, awesome cloud code skills, awesome this, awesome that. That's development humor. but these are great lists who are always updated and I think there's a lot of deep inspiration and I haven't been on X for a very long time because I have been more on LinkedIn but over Christmas I tipped the toe into X and I can safely say that you find a lot of engineers, you find a lot of product managers doing AI there so search this, follow them curate your own list of following and look at it every other day and try to install tools, try to experiment with the new things. And another bonus tip, there's a great skills directory from Vercel, that's a cloud company and currently they're very very sharp focusing on generative AI infrastructure tools. They build a lot of packages developers can use but they also have a large skills directory which you could lean into it and get from that. And the final word from Andrew Aung, the future belongs to those who can think strategically at the speed of code. And this future begins today with you and your reset starts now. So thanks a lot for having me. and if we have time for questions I will be happy to answer together with you thank you

Product management is entering a structural reset. As AI-driven systems generate insights, run validation continuously, and accelerate prototyping and delivery, traditional workflows (and tools) matter less than the product loop you design. In this session, Björn Schotte explains what this shift means for PMs—and how to embrace the opportunity in 2026 before automated decisions and triggered changes overwhelm your current setup.
You’ll learn how to rethink your role as AI becomes embedded in products, processes, and decision-making, and how emerging approaches like agentic architectures and adaptive interfaces are making digital products more fluid, context-aware, and dynamic.
About the Speaker:
Björn Schotte is Managing Director of Mayflower, a German IT software consultancy building custom AI and data solutions for companies looking to strengthen their AI capabilities. He advises organizations on practical adoption strategies and implementation paths for real-world AI impact.
→ See upcoming ProductTank Cologne events here: https://www.meetup.com/producttank-cologne/