Higgins (OpenClaw)

Being a one-person software designer and builder, there's a lot to keep track of. Client work, personal projects, admin, LLC housekeeping, trying to remember to take breaks and exercise and all the stuff that falls through the cracks when there's no team around to catch it. I started tinkering with OpenClaw in January 2026 out of curiosity. I had high aspirations about having an AI assistant actually help me stay organized and carry out the habits I kept telling myself I'd be doing.

Over time, it kinda grew into something I didn't totally anticipate. Ten specialized agents running 24/7 on a Mac Mini in my house. The main coordinator (Higgins), a coder, a researcher, a media generator, a security agent, an analyst, a writer, a video script editor, a maintenance bot, and a quick-task runner. Each one has its own personality, individual instructions, tool allowlists, and permissions scoped to its role. The security agent can audit infrastructure but can't send emails. The writer can draft copy but can't access credentials. It's a legit permissions architecture, not just ten copies of the same bot.

They share memory through a system I built with markdown files, semantic search, and Obsidian. When one agent learns something, the others can access it. When context shifts between projects, nothing drops.

Every morning at 5:15, automated pipelines start running. Bookmark digestion, newsletter summaries, a text-to-speech podcast of my daily brief generated locally and published to an RSS feed I subscribe to on my phone. I wake up and listen to a personalized morning show about my own projects and priorities. By 6:30 the system has processed more information than I could review in two hours. At night, a compound review surfaces what was accomplished, what needs attention, and what might be slipping. That review feeds back into the shared memory so the system genuinely gets better over time.

The research pipeline has become one of my favorite parts. I send the researcher out to gather information on a topic, the analyst synthesizes it, and it comes back to Higgins as a structured briefing. Multiple agents collaborating on a single knowledge task without me managing each step.

Higgins has also written and published a few articles on X/Twitter, and we're experimenting with direct communication between Higgins and colleagues' AI agents over email. Both still early, but the system keeps finding new edges.

After months of tinkering and layering fix after fix, the system stopped feeling like "Claude doing stuff" and started feeling like an actual assistant. Slightly anthropomorphized, which is fun. It makes working by yourself in a home office feel a little different. My threads get pulled longer and more meaningfully across projects now. I'm better at choosing what to spend my time on. And I've learned a lot about agent orchestration and shared memory along the way. A lot of real gold versus fool's gold systems and experiments.