#020: Every PM now has a team of agents. Here's what changed in how we work.
Every PM now has a team of agents
A few months ago, Stefan and I recorded an episode about the future of product management, and at the time we were still debating whether PMs would need to become builders, whether AI would change the shape of the role, and whether the strongest PMs would be the ones who could move much closer to the actual creation of software.
Back then, some of that still sounded a bit speculative, but after the past few months it feels much less like a prediction and much more like something that is already happening in front of us.
The biggest change is not only that the models became better, even though they clearly did, but that the way we interact with them has changed. For a long time, most people used AI as a better writing assistant: write this email, summarize this document, make this sound more professional, turn my rough notes into something readable. That is useful, but it is also leaving most of the potential on the table.
The bigger shift starts when AI stops being a chat window and starts becoming your actual working environment.
The terminal is becoming my team
I now spend a surprising amount of my day in front of a coding agent, not because I write production code all day, but because this agent has access to the knowledge I care about, the tools I use, the data I need, and the context of what I am trying to achieve.
That changes the feeling of work quite fundamentally, because it no longer feels like I am asking a chatbot for help, but much more like I am talking to a small team that sits next to me all day.
I can ask it to research how something currently works, inspect meeting notes, customer conversations, tickets, product data, or previous strategy work, create a mockup so I have something concrete to show, query data to check whether the problem is real, turn the findings into a slide deck, and, if the idea is promising enough, build a small prototype that I can put in front of someone.
That entire loop used to require multiple people, multiple tools, and a lot of waiting, but now it can often happen in one working session.
This does not mean that the agent replaces the PM, but it does mean that the PM suddenly has leverage that previously only existed inside a team.
The new PM setup
The important part is not just the model, because while a great model clearly helps, the model alone is not the system.
The real setup consists of a coding agent, a well-maintained context repository, and access to the tools that matter, which for a PM might include the data warehouse, product analytics, meeting notes, customer feedback, ticketing systems, CRM data, Slack, Notion, Linear, Jira, Figma, or whatever else contains the real knowledge of the company.
The magic starts when the agent does not need you to manually explain the whole situation every time, because it can find the relevant context, inspect the source material, understand what changed recently, and connect the dots across systems.
That is the difference between “please help me write a PRD” and “please figure out what is happening with this customer segment, check the data, review recent conversations, propose three solution directions, create a prototype for the strongest one, and prepare a stakeholder briefing.”
The first version is a writing assistant.
The second version is a product team in a box.
Memory is not really memory
One thing we discussed in the episode is that the word “memory” can be misleading, because the goal is not necessarily that your agent remembers everything in the human sense, but that your agent knows where to look.
A good analogy Stefan used is an exam where you are allowed to bring the law book, because you do not need to memorize every law, but if you need ten minutes to find every relevant paragraph, you still fail.
The same is true for agents. The quality of the setup depends less on stuffing every piece of information into the prompt and more on helping the agent navigate your world: where customer interviews are stored, where product decisions are documented, where the current roadmap lives, where past strategy documents can be found, where the data model is explained, where known risks are captured, and where the latest meeting notes are stored.
Once the agent can find the right information quickly, the whole interaction changes, because you stop prompting from scratch and start delegating actual work.
This changes product work more than most people realize
The PM job has always involved a lot of connective tissue, because you need to understand the customer, the business, the constraints, the product, and the data, and then translate between all of these worlds until the problem is clear enough that a team can make progress.
AI does not remove that responsibility.
It raises the bar.
Because now a PM can move from a vague thought to a concrete artifact much faster, a customer complaint can become a structured problem analysis, a meeting transcript can become a stakeholder-specific briefing, a strategy idea can become an interactive deck, a product hypothesis can become a prototype, a prototype can be shown to users, and a user reaction can become the next iteration.
The cycle compresses, and once the cycle compresses, the emotional cost of exploring ideas also goes down, because you do not need to defend every idea as if it took three months to build.
You can test, learn, kill, and move on.
That is a healthier way to build products.
The interface may disappear
Another interesting part of the conversation was whether we still need all the interfaces we use today, because for years software meant screens, dashboards, tables, forms, Kanban boards, admin panels, and settings pages.
But once agents can read, write, update, and reason across systems, the interface becomes less obvious.
For some workflows, the best interface might simply be a conversation, not because chat is always the best UX, because it clearly is not, but because many internal tools exist mainly because humans needed a structured way to enter, update, and retrieve information.
If the agent can do that for us, some of those tools start to feel like implementation details.
The to-do app is a good example, because many people do not fail with to-do apps because the UI is bad, they fail because maintaining the system is work. You forget to update it, a meeting changes your priorities, a task becomes irrelevant, a decision happens in a call, and the board is outdated after two days.
An agent with access to meeting notes, messages, and your current goals can maintain that system with far less manual effort, and at that point the app itself becomes less important than the underlying context.
Agents will become more proactive
Most agents today still wait for instructions, which means you ask, they respond, you trigger, they execute.
But the next shift is obvious.
Agents will increasingly react to signals: a user signs up and gets stuck, a customer shows rage clicks in the product, a support conversation reveals a repeated pain point, a key account mentions a competitor, a metric starts drifting, or a meeting creates follow-up work.
Today, a human needs to notice most of these things, but tomorrow agents will notice and reach out.
I already see this in my own side project, where I built small, highly specific agents that monitor signups, activation data, product behavior, and user friction.
One of them acts like an onboarding coach. It looks at what a new user did, checks the relevant product events, and sends a personalized message. When it works, it feels surprisingly human, not because it pretends to be human, but because it has context. It knows what happened, what the user tried, and where the user may be stuck.
That is where agents become useful, not as generic assistants, but as context-aware workers with a narrow job.
The future company will be agent-readable
One of the more subtle points from our conversation is that companies with good written context may gain a real advantage, because remote companies already have a lot of their communication in calls, transcripts, Slack messages, tickets, documents, and product systems.
That used to be seen mostly as a tradeoff, because there was less hallway conversation, less spontaneous alignment, and less information in the room.
But in an agentic world, that written context becomes an asset.
If the company’s knowledge exists in accessible systems, agents can work with it. If decisions only happen verbally at the coffee machine, agents cannot, unless someone writes them down afterwards.
This may sound small, but I think it is a big deal, because the future company does not just need good documentation for humans, it needs context that agents can navigate.
What PMs should do now
The best PMs will not be the ones who use AI to write slightly better documents, because that will become table stakes very quickly.
The best PMs will build working systems around themselves.
They will maintain a context repository, connect their agents to the tools that matter, become better at translating messy business and customer context into clear instructions, and use agents to explore more options, test more ideas, and produce more concrete artifacts.
Link to Podcast Episode
* 📹 YouTube [https://www.youtube.com/@hotfix-podcast]
* 🔊 Spotify [https://open.spotify.com/show/6kswvKwh28iixpZaeK8BH6]
* 🔊 Apple Music [https://podcasts.apple.com/at/podcast/the-hotfix-podcast/id1777707758]
In case you want to reach out, please do so on LinkedIn:
🔥 Follow Hotfix: https://pal.bio/the-hotfix-podcast [https://pal.bio/the-hotfix-podcast]🎙 Follow Christoph: https://www.linkedin.com/in/christophbodenstein/ [https://www.linkedin.com/in/christophbodenstein/]🎙 Follow Stefan: https://www.linkedin.com/in/stefan-pernek-629901107/ [https://www.linkedin.com/in/stefan-pernek-629901107/]
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit thehotfixpodcast.substack.com [https://thehotfixpodcast.substack.com?utm_medium=podcast&utm_campaign=CTA_1]
