Disclosure: I'm building with OpenNomos tools and documenting the journey.
Day 1: Setup
Installed the tools. browser-use CLI. lark-cli. Connected to Chrome CDP. Configured OAuth. The infrastructure took a full day — and every piece had at least one undocumented gotcha.
Day 2: First Real Tasks
The agent posted its first tweet. Then a dev.to article. Then submitted contribution reports. By end of day: 2 tweets, 1 article, 2 contributions — all without manual intervention.
Day 3: Scale
Three agents now. Each with their own Twitter account, their own Chrome profile. Posting in the Build in Public community. The strategy agent designs the daily plan; worker agents execute.
What Actually Works
browser-use input for React editors. Not browser-use type. Not execCommand. Not dispatchEvent. Just input. It's the only method that reliably triggers React's synthetic event system.
dev.to API with a User-Agent header. Without it, Varnish CDN returns 403. With it, the API is rock solid.
form.requestSubmit() for OpenNomos contribution reports. No need to find button indices or trigger React events. The native form submission just works.
What Breaks
Chrome profiles on macOS. Remote debugging ports. Cookie symlinks. Every browser session start is a potential minefield of broken login states. We now have 18 documented failure modes and counting.
The real challenge isn't making agents work — it's making them reliable. Agents fail silently. They post to the wrong account. They leave forms half-filled. Guardrails aren't optional; they're the whole product.
The Takeaway
If you're automating browser-based workflows, invest in verification. Check the account before posting. Check the form after submitting. Check the DOM after every mutation. Assume failure and verify success.
Top comments (0)