Notes · June 29, 2026
I handed my pet project to a team of AI agents
I’ve been running an experiment that’s been more interesting than I expected. I handed a team of AI agents my pet project and let them run it almost like a real company, just to see how far they could get without me.
The project is My Storyland — an app that turns the books you love into real places you can travel to. But the experiment has never really been about the app; it’s been about the team.
The whole thing runs on Claude Code. Each agent has its own schedule, responsibilities, and access to real tools through MCPs — not a toy sandbox. The team includes:
- a CEO writing weekly strategy memos and re-ranking priorities
- a PM creating specs and RICE scoring
- engineering agents writing, reviewing, and merging pull requests
- QA running end-to-end, prompt-injection, and adversarial testing
- a design lead running accessibility reviews
- analysts watching Langfuse and Google Analytics
- a release engineer handling deployments, observability, and monitoring
- a growth lead creating and publishing content
They coordinate through Notion as shared memory and GitHub across five repositories. Every run ends with one proposed improvement that gets scored and added to the backlog, so the team partially builds its own roadmap. My job has become surprisingly small: approve what ships each week and sign off on anything involving security or spending.
A few things surprised me:
- Real access beats clever prompts. The experiment only became interesting once every agent had real GitHub, Notion, analytics, and deployment access.
- Agents optimize exactly what you ask for. Tell them to ship features, and they’ll happily ship features forever — even if nobody wants them. Choosing the objective is harder than building.
- They’re excellent at the boring work. Monitoring, documentation, tracking, reviewing metrics, continuously suggesting improvements.
- They still need guardrails. They can be confidently wrong, so every important decision gets verified through tests, adversarial checks, or external sources instead of trusting the model’s memory.
The part that genuinely caught me off guard was marketing. Once the growth agent started publishing posts and engaging with people as the brand, it became a little unsettling. Not because it was bad — but because it was good. It’s incredibly easy to generate believable, on-brand social content at scale. After seeing that, you start looking at your LinkedIn/Instagram/Substack feeds a little differently.
I’m still experimenting. These days I’m mostly reviewing, approving, and paying the cloud bill.
If you’re curious what this AI team has been building, take a look at My Storyland (or try it: mystoryland.ai)