“Engineers are 10x faster. Product managers haven’t sped up at the same rate. Now they’re the bottleneck.”
I came across this Andrew Ng quote shared by Lenny Rachitsky, and it stuck with me. AI coding agents are already changing how fast software teams can create code. But what happens when the bottleneck moves to the people who define requirements, evaluate options, prepare decisions, and carry domain knowledge?
That is the work many product managers actually do.
So I started a test: I treated knowledge and information like code, and I used AI agents not to write software, but to manage my own knowledge work.
The bottleneck moved, and it’s us#
I’m a product manager. I don’t ship production code. I sit in the middle of partner strategy, customer requirements, ecosystem decisions, and communication. More than 110 partners in the ctrlX OS ecosystem, plus a steady stream of new companies asking to join, new use cases to evaluate, and strategic decisions to prepare.
AI coding agents like Claude Code and Codex have made writing code cheap. But that doesn’t remove the bottleneck - it moves it. Someone still has to decide what gets built, which partner fits, why one option beats another. That work, the knowledge work, has not become faster.
So I asked a simple question: if these agents accelerate engineers, why not me?
You don’t need to be a developer to do this. If you can write a structured document, you can start. Let me walk you through it.
Lesson 1: For knowledge workers, “code” is Markdown#
The name Claude Code is misleading. These agents aren’t limited to programming. The moment you understand what plays the role of “code” in knowledge work, they open up completely.
That role is filled by Markdown. It is a plain and simple text format that every large language model reads exceptionally well. When I keep my knowledge in Markdown, the agent can read it, condense it, update it. And it treats every change like code: a clear diff that I review and approve or reject. Nothing happens behind my back.
I have been using Markdown for my notes since 2021 with tools like Logseq and Obsidian. The new part is not Markdown itself. The new part is that an AI agent can now work directly inside this knowledge base.
This sounds unremarkable. Its consequences are not.
Lesson 2: Climb the ladder from chat to knowledge base#
I went through four stages. You can recognize which one you’re on right now.
Stage 1: Chat. Paste a prompt from your prompt collection file, adjust it, add a company name, and copy the answer from the chatbot into your document or presentation by hand. This is how most people have worked since ChatGPT launched. It works, but it’s manual and nothing accumulates.
Stage 2: An agent inside the chat. Save that long prompt once as a reusable agent or create a project. Now you only enter the company name. Much better, but the result is still trapped in the chat window.
Stage 3: An agent that writes files. The agent runs inside your editor and produces real Markdown files. It creates profiles, organizes folders, writes straight into your knowledge base. The output finally leaves the chat. No copy and paste of information.
Stage 4: The LLM Wiki managed by the agent. A git-versioned knowledge base, following Andrej Karpathy’s concept. One central file tells the agent what the wiki is for and what it can do. Every recurring task, such as a company analysis or a meeting summary, sits beside it as a reusable “skill.” The structure itself tells the agent what to do.
The leap that matters is from stage 2 to stage 3: the moment your results stop being throwaway chat messages and start becoming a knowledge base that compounds.
Lesson 3: Start with the work you already repeat#
Don’t begin with a grand vision. Begin with the task you do over and over. Mine were obvious:
I no longer accept partner meetings without a transcript. That sounds strict, but it’s the foundation of everything: from each transcript, the agent drafts a partner profile, open points, next steps, and frequently asked questions for our partner handbook. Knowledge that used to evaporate now becomes reusable.
When a sales colleague needed arguments for a specific new customer, I pulled together market research, competitor observations and customer context. The agent turned it into a structured briefing: customer situation, likely objections, relevant ecosystem examples, and suggested talking points. It was not the final answer, but it was a much better starting point than a blank slide.
And company analyses for partnership decisions: company name in, structured report out: key figures, pros and cons, decision-makers, portfolio fit, and proximity to competitors.
What still needs human judgment#
This is not magic, and it is not autopilot.
Transcripts can be wrong. Agents can miss context or make a connection that sounds plausible but is not true. Confidential information needs clear boundaries. And every recommendation still needs a human review.
For me, the point is not to outsource judgment. The point is to stop wasting time on collecting, copying, and reformatting information. I want the agent to prepare the material so I can spend more time on the decision.
Was the path frictionless? Not entirely. I deliberately do all of this with corporate-approved tools. We use Gemini and Microsoft Copilot at work. I use Gemini as an agent inside VS Code. It is less mature than Claude Code or Codex, which I use for my private projects. Sometimes it does not write its answer to the Markdown file. Sometimes it will not create folders on its own. The approach still works. And that’s the whole point: this is reachable today, not in two years, and not only with the most advanced tool.
The question I’m sitting with#
I’ve been running this knowledge base for three months. It grows with every meeting, every piece of research, every decision. Which brings me to the question I find most exciting:
What happens when not just one person but an entire team works this way? On a shared knowledge base, with branches and pull requests, the way engineers have handled code for years? Could this become the next generation of how a company manages its knowledge?
I don’t have a finished answer. But I’m convinced of one thing: you only understand AI by doing it.
Pick one recurring task this week: a meeting summary, a customer briefing, a partner profile, a research note, or a decision memo. Put the input into Markdown. Let an agent create the first version. Then review the diff.
That is where it starts.
👉 Are you already using AI agents for knowledge work or only for code? Let’s talk on LinkedIn
