What I've learned shipping Rapid Claw with AI agents (the messy, honest version)

Hey IH 👋

Been heads-down building Rapid Claw (https://rapidclaw.dev) for the last few months and figured it was time to surface and share some of the messy stuff. Not a "we hit $X MRR" post — more of a "here's what nobody told me about building with AI agents" post.

Quick context: Rapid Claw is a project I'm building mostly by orchestrating AI agents instead of writing every line by hand. I went in fairly skeptical and came out converted, but not in the way I expected.

A few honest takeaways from the journey so far:

1. Agents crush the boring 80%, and confidently faceplant on the spicy 20%
   The plumbing — CRUD, schemas, tests, refactors, glue code — ships faster than anything I've built before. But the moment a decision involves taste, tradeoffs, or weird infra (auth flows, queue strategy, billing edge cases), the agents will sprint off a cliff while smiling. I now treat agent output like a junior dev's PR: ship the easy stuff, redline the hard stuff.

2. Infra is where I bled the most
   The "build" was fast. The "make it not fall over at 2am" was not. Observability, retries, idempotency, model provider rate limits, cost guardrails — that whole boring layer ate way more hours than the actual features. If you're starting a new agent-powered product, budget at least 30% of your time here from day one. Trust me.

3. Distribution still > features
   Classic, but it hits differently when you can ship a feature in an afternoon. The bottleneck was never code — it was finding the people who actually had the problem. Posts like this one have moved the needle more for me than half the features I've built.

4. The first paying users completely rewired my brain
   I'm not going to flex revenue numbers because it's still small and lumpy, but the first real customers changed my prioritization overnight. "What's cool to build" got replaced by "what does the human who already paid me actually need." That single shift was worth more than any framework or productivity hack.

5. Working with agents is a skill, not a feature
   The leverage comes from knowing how to scope work, set up the right context, and verify output fast. The folks I see struggling tend to be the ones treating agents like a vending machine. The ones winning treat them like a team you have to manage.

If you're building something similar (or thinking about it), I'd love to compare notes — what's your stack, where have agents surprised you (good or bad), and what's been the hardest part of shipping?

And if you want to peek at what we're building → https://rapidclaw.dev — feedback always welcome, the more brutal the better.

Happy to answer anything in the comments.

—Tijo