A personal log of what happens when someone decides to take AI seriously — from first hello to running a multi-agent research system.
What drove me to build my own AI system when Claude, ChatGPT, and Gemini already existed. The research, the hardware wait, two days of failed installs, a full Mac wipe, and the moment it finally clicked.
The Mac Mini arrived. Two days of failed installs, a full wipe, a fresh start — and the moment it finally clicked that the prompts weren't the point. The connections were.
If there's one agent, can there be more? Yes. A lot more. How I discovered research swarms, decided I wanted five agents, and built a morning briefing pipeline in a single session.
How switching from general web search to Tavily cleaned up the research pipeline — and what makes an AI research swarm fundamentally different from a news app.
I asked my home AI for ideas beyond the morning briefings. It suggested email drafts and spreadsheets. I had a better idea: stop paying GoDaddy. Two sites migrated to Cloudflare Pages, one friend’s business site redesigned at a swim meet, and a cancelled hosting plan.
I watched a video about agents arguing about stocks. Then I realized the same pattern works for hiring decisions, purchases, architecture choices — anywhere you need a second opinion that actually pushes back.
The evolution from "pre-input agent" (rejected) to one Gemini checker to two checkers to a full agent council — and what the conversation that got there actually looked like.
Claude + Gemini 2.5 Flash + GPT-4o-mini, in parallel, before every reply. Here's the architecture, the cost breakdown (~$0.0003 per check), and why I decided the latency was worth it.
The original research swarm had two overlapping layers and a blind spot on world news. Here's the redesign: three dedicated researchers, a bull/bear debate before every picks run, and a timing split that eliminates redundant fetches.
I asked my home AI to generate a logo using GPT Image 1. Back and forth on colors, spacing, background. Uploaded it to the site — and it was tiny. The real lesson: one line of CSS was the right answer all along.
I saw a video about Hermes and its self-improvement loop. Could my AI agent do something similar? I asked — and it proposed the architecture, explained the tradeoffs, and built it when I said yes.