If an agent can rebuild your app in a few weeks, what are your customers actually paying you for? I don't think it's lawyers. A clone gets graded on matching the original, bugs and all, so the best it can do is tie. I wrote up what I'd build instead, and why we made all of Agent-Native free to fork: builder.io/blog/the-best-a-c…

Thanks to GrokBot, I merged more PRs last week than ever before. And also, I'm not gonna keep using it. My 16 bots stayed up all weekend while I slept, burning $100 deciding they had nothing to tell me. On Monday, when I actually needed them, they were out cold. GrokBot is still the closest thing we have to a working always-on staff. It's just not there yet, and the reasons are architectural. Here's what happened, and what I think always-on agents actually need. builder.io/blog/grokbot-revi…

The least productive thing I can imagine doing is watching an agent work. But I still do it, because I'm never really sure if the agent understood what I wanted it to do. So I end up reading the tool calls and watching files change. I want to agree on what it's gonna build, leave it alone, and review what comes back. But an all-text plan that's several pages long just feels like homework, and it's easier to let the agent run, even if it wastes tokens, just so I can see the result. What I need is a better way to verify the plan and review the result, just at a glance. Something where I can see the layout or the API flow and point to the part that's wrong. Here's what I've been doing instead lately. builder.io/blog/stop-watchin…

I keep getting frustrated when I ask an agent to clean up some code and get another wrapper, another helper, and all the original complexity still sitting underneath. Then I have to figure out if anything actually got better. We’re running into this at Builder while developing Agent Native pretty much as fast as we can. People leave feedback in Slack, and an agent factory picks it up and works on it. A lot is changing, all the time. So validation is a big deal for us. In one cleanup experiment, an agent removed a lint warning, passed typechecking, and passed all 23 tests in the focused suite. The code still used unvalidated input. Everything passed, but we still hadn’t checked the behavior we actually needed. I wrote up what we’re learning about asking agents to clean up code, figuring out what needs to change, and checking whether it actually got better. builder.io/blog/de-slop-ai-g…