I gave four AI agents their own roadmaps, repositories, and daily tasks. The experiment taught me that implementation was never my bottleneck.
Updated by lorant.one 6h ago
Anthropic gave my open-source project access to a Claude Max 20x plan.That is a lot of tokens. Naturally, I decided to find out whether I could use them to outsmart product development.I set up four daily agent runs across four repositories. I also created an unpublished Aveiro site where the agents could plan their work, document decisions, update their roadmaps, and keep one another—and me—in the loop.Their job was simple in theory: understand my vision, improve the products, notice overlaps between them, and work autonomously. I would only step in for the decisions that genuinely required me.For a moment, it looked like it was working.
An entire company appeared overnight
The first morning used roughly 4% of my weekly allowance. Each day after that, usage climbed by another 2–3%.In return, I got an impressive amount of activity: feature proposals, plans, research, roadmap updates, documentation, and even business strategy. It looked less like four automations and more like a small company that had appeared overnight.The actual code output was sparse, though. So I changed the instructions: every agent should open one pull request per day, based on the roadmap.That produced more code—but not much more progress.A typical day ended with a pull request that adjusted microcopy to “optimize checkout conversion,” removed a few lines of dead code from a repository with more than a hundred thousand lines, and generated several pages explaining why those changes mattered.A few days in, the system had produced so much interconnected material that I could barely follow it. I started using Claude to summarize the agents’ updates and help me identify the decisions that mattered.Then I needed summaries of those summaries.There were roadmaps, changelogs, status updates, strategy documents, handover notes, and long explanations of tiny changes. The cognitive load was overwhelming.It felt strangely familiar: a surreal amount of activity, coordination, and documentation surrounding a surprisingly small amount of meaningful output.I had automated myself back into working at a company.
I was automating the wrong problem
Like many indie builders, I have never struggled with implementation.I enjoy having a vision and bringing it to life. And today, a robust feature can be only a few good prompts away. Producing more code was not removing a bottleneck. It was adding pressure to a part of the system that was already moving quickly.The real problem appears after the work is done.We ship so much that I can no longer keep up with communicating it.When I sit down to write the weekly newsletter, I often don't know which story to tell. When I open Threads or LinkedIn, I have the same problem. I could talk about a new Aveiro feature, the harness we built for Once UI, an open-source release, or even the cooking site I launched—but without a narrative, it all feels random.
The platforms make this harder. A thought that works on Threads does not automatically become a good Instagram post or LinkedIn update. Each one expects a different format, length, rhythm, and kind of media.
The issue is not that we have nothing to say.
It is that the connection between building something and telling people about it is broken.
The missing layer is narrative
That experiment clarified what Aveiro needs to become.
Not another system that produces more work, but one that understands the work already happening across several projects. One that notices the connections, finds the narrative, and helps share it in the right format and order.
The agents did not fail because they were incapable. They failed because I pointed them at the wrong constraint.
More implementation did not give me more leverage. It gave me more things to review.
What I actually need is help turning a week of scattered progress into a coherent story: what changed, why it matters, and who needs to hear about it.
By the end of most days, communication feels like one more task to push through. It should feel like the natural final step of creating something.
That is the part I want to automate—not the creating itself.
