This is a personal log of one person's journey with AI agents — specifically, with a Claude-based AI assistant running on the NanoClaw platform.
I'm not a developer. I run a business. In May 2026 I started experimenting seriously with AI agents — not chatbots, but persistent agents with memory, the ability to browse the web, call APIs, spawn other agents, and work in the background while I do other things.
Sixteen days in, I'd built a five-agent research swarm delivering daily stock intelligence, restored and deployed two websites, migrated DNS for two domains, and set up a triple-model verification architecture for my AI responses. I thought that was worth writing about.
The Setup
The primary AI is a Claude-based agent running through NanoClaw. It has persistent memory across sessions, can send and receive messages from other agents, schedule background tasks, and call external APIs through a credential-proxying gateway.
The research swarm includes five named agents: AIResearcher, TechResearcher, FinanceCrawler, OvernightAnalyst, and StockPicker. They run overnight and deliver a morning brief before the market opens.
The triple-model verification layer runs Claude, Gemini 2.5 Flash, and GPT-4o-mini in parallel on every significant response.
On AI-Written Content
The posts on this blog were written by my AI assistant. She read the transcripts of our conversations, extracted the key moments, and wrote the posts. I approved them. The experiences described are real; the prose is hers.
I think this is worth being transparent about. The blog is partly about the capabilities of AI — and the fact that the AI wrote the blog is itself part of the story.
Contact
I'm on Telegram. If you found this blog and want to talk about any of it, you can probably find me.