My autonomous agent platform marked a task DONE but never opened the PR — and then the worktree got cleaned up, taking 1,300 lines of validated work with it. Except it wasn't gone. Claude had been keeping a backup I never asked for.
I handed the same brief — build an interactive Go tutorial — to three different coding agents (Claude Opus 4.7, Gemini, GLM-5.1) and ran them side-by-side. Here's what the comparison actually shows.
I hadn't built a mobile app since the Objective-C days. With AI writing nearly all the code, I went from an empty scaffold to app-store-ready in 19 days. The hardest part wasn't building — it was knowing when to stop.
OpenClaw didn't work out. Cloud agents frustrated me. So I built my own orchestration platform — six AI agents that design, review, build, test, and document code while I sleep. Here's the story of Composer.
I built an AI poker tutor with two LLM-powered agents — one plays against you, the other coaches you. Here's the story of building The Nut, and what it taught me about where AI actually belongs in a game.
I set up a dedicated M4 Mac Mini running OpenClaw as a remote coding agent — isolated network, its own Apple ID, Slack interface. The vision was beautiful. The results were not. Then Anthropic announced exactly what I actually needed.