AI Founder Warns: Project Managers Must Embrace Agentic Software
Henrik Kniberg shares how AI agents are reshaping software development, from automated releases to autonomous backlog management. He argues that project managers who fail to learn how to work with AI will quickly become the bottleneck.
AI Impact: AI is revolutionizing software development, allowing teams to deliver higher quality work more efficiently.
Project Management Shift: Project managers must adapt to integrate AI, or risk becoming bottlenecks in the development process.
Agile Practices: Maintaining an agile approach is crucial; minimal bureaucracy enables faster project delivery with AI's assistance.
Agent Utilization: AI agents handle tasks like release management and backlog organization, freeing teams to focus on complex issues.
Future Skills: Learning effective AI context management is vital for project managers to stay competitive in a changing landscape.
Henrik Kniberg has a background in project management at Spotify, LEGO, and Mojang. Currently, he is a cofounder and chief scientist at Abundly.ai. And he's the author of "Generative AI in a Nutshell".
We sat down with him to discuss how AI is changing software development — and by extension, project management. He said project managers need to get on board or become the bottleneck.
From product to AI startup
I am the cofounder and chief scientist at Abundly.ai, where I lead the development of our platform for autonomous AI agents.
Unlock for Free
Create a free account to finish this piece and join a community of forward-thinking leaders unlocking tools, playbooks, and insights for thriving in the age of AI.
Have an account? Log In
I previously worked as an Agile coach at Spotify and LEGO, and I did Minecraft development at Mojang. In fact, I made a video called "Spotify Engineering Culture," which people often refer to as the "Spotify Model," describing their product development approach.
And I'm the author of "Generative AI in a Nutshell".
How AI lets software teams ship more with less
Software development can now be done so much faster with AI help, and that's changing project delivery dramatically.
It lets me and my team focus more on architecture, design, and UX — where we should be spending our time — and spend less time typing and debugging lines of code.
The result is that we can deliver more, with higher quality. And we do it in less time with a smaller team.
With AI, we can deliver more, with higher quality. And we do it in less time with a smaller team.
Why teams should push for minimum viable bureaucracy
We aren't moving away from traditional project management methods — we never had them in the first place.
We started off using agile methods, and we're working hard to keep an agile approach even as we grow. We have a one-week cadence, where we sync on Fridays to go through progress and priorities. We maintain a simple Kanban board with WIP limits and a triage column that works like an inbox that gets cleared every week — each item goes either in or out.
Every commit is deployed automatically to a shared test environment. And we push to Production about twice per week, or more often when needed.
We also do an informal sync most days. We just get together and talk about what we are doing that day.
Overall, we try to keep a minimum viable bureaucracy. And AI is helping us do that.
We try to keep a minimum viable bureaucracy. And AI is helping us do that.
How agents are transforming release and change management
We experiment with agentic workflows every day, as it is our core business.
For example, we built a release manager agent. Whenever we want to release to Production, which happens about twice per week, we ping that agent, either in Slack or in our app.
When pinged, the agent checks the latest changes on our test branch, analyzes the commits, and generates and posts two versions of release notes — one for internal use and one for external use that omits changes that aren't customer-facing. It then analyzes our documentation site to see if anything needs to be updated and makes a Pull Request for any changes.
We are constantly tweaking the instructions of this agent. The latest tweak was checking for security issues.
Another example that we're working on now is an agent that handles internal product change requests. When a stakeholder reports a problem or requests a feature, the agent automatically evaluates the priority, examines the code to check the complexity of the change, and posts an analysis on Slack. It then determines whether it should autonomously fix the issue and make a PR, or if it is uncertain and should therefore ask us first. It also creates a ticket in our backlog to track progress. And when it makes a change on its own, it tests the change in a browser to evaluate UI impact, adds screenshots to the PR, and reviews the code quality.
The purpose of that agent is to automatically handle all the simpler tasks, so my team can focus on the more complex tasks. Though, to be fair, we also do those complex tasks with the help of AI!
Our other use cases include research, screening, compliance auditing, document routing, and more.
Why humans need to stay in the loop with AI-driven engineering
So, we are basically trying to make AI a part of every process. For example, AI in backlog management can implement some tickets, and do release management.
We are basically trying to make AI a part of every process. But we keep a human-in-the-loop where appropriate.
But importantly, we keep a human-in-the-loop where appropriate.
Although we use AI for development, it is more like a pair-programming partner than outsourcing to some offshore company. That is, we are still deeply involved in the code — we just let AI write most of it.
How an AI-first engineering stack works
Speaking of development, we use Cursor with Claude Opus. And we are experimenting with Claude Code as well.
Cursor + Claude is an incredibly powerful combo, like pair-programming with a blazing-fast genius colleague. And we've created a quite comprehensive set of rule documents to help provide context to the AI model.
Our product is deployed in Google Cloud Run and Vercel, with continuous delivery, so we put stuff into Production by simply making a PR and merging to Main — or asking an agent to do it. Our code is Typescript, both frontend and backend.
Having the same language throughout the entire stack makes life easier for both our agents and us.
Having the same language throughout the entire stack makes life easier for both us and our agents.
All our code is in a monorepo, meaning that distinct projects, apps, and libraries are in the same repository. That also simplifies things, since a single commit can contain both frontend and backend changes. It makes dependency management easier.
As far as project management, we use a Kanban board in Notion to organize and visualize our work and priorities. And we set up an agent to interact with it. Here's a typical scenario: Someone brings up an issue on Slack, and after a few messages back and forth, someone writes, "Hey @backlogger pls add a ticket for this". The agent interprets the Slack thread, looks for similar tickets on our Kanban board, and either updates an existing ticket or creates a new one, attaching relevant context and screenshots from Slack. Super useful!
We use this agent every day. Our planning meetings are a lot faster now because every ticket has a super clear description, screenshot, and link to the Slack thread. In the future, we're going to extend this so it also suggests priorities and fixes simple issues automatically.
How Claude Opus 4.5 changed the game
Overall, I'm constantly surprised by how fast the tools and models are evolving and improving. It forces us to continuously evaluate the latest.
Recently, Claude Opus 4.5 was released, and it is really good at doing more complex code changes. Because of that, we are now more likely to "one-shot" a problem — i.e., write one detailed prompt — rather than work in small increments with human involvement at every step.
Basically, we ask it to solve a tricky problem while we go have lunch.
It is surprisingly effective. We still need to review the code, but most often, it is fine.
Why project managers must learn to work with AI agents
Moving forward, project managers will need to get used to having a team consisting of human and AI colleagues working together. That means you.
You, as the project manager, need to get used to working with AI agents as colleagues. If you don't, you will be the bottleneck in the process. Everything will move at a pace that you can't keep up with.
Not knowing how to use AI effectively will be like not knowing how to use the internet effectively. Sure, you can still do project management without using the internet. But you'll be severely handicapped, and your job prospects will be bleak.
Not knowing how to use AI effectively will be like not knowing how to use the internet effectively.
How better context management unlocks more reliable results from AI
The field keeps changing, so the key thing is to experiment.
Experiment with different AI tools and use cases. Build your skill as a user and context engineer. And look at what you and your team spend time on — where can AI help save time or improve quality?
Beyond experimentation, context management — understanding how LLMs work with context — is crucial. If you study that, you will be able to choose and use AI tools more effectively. Most AI fails happen because of bad context management. In other words, it fails because it doesn't have the information it needs.
Nobody knows what's going to happen in the future, but if you are good at using generative AI, you will be in a better position to deal with whatever the future holds.
Follow along
You can follow along as Henrik Kniberg pushes agentic boundaries in PM on YouTube, LinkedIn, and X. And don't miss his product, Abundly.ai.
More expert interviews to come on The Digital Project Manager!
