Dolphin is a multi-plane AI agent system running on my own business since early 2025. Three machines, 75+ tools, persistent memory across three knowledge graph databases, governance layers, SMS interface, and a build pipeline that deploys to a Windows host over SSH. It is currently running. I am probably not looking at it.
It does my CRM, my call transcription, my client dashboards, my prospecting, my analytics, my scheduled work, and overnight autonomous tasks. I didn’t build it to show anyone. I built it because I needed it. Then I started selling the same architecture to clients, because apparently other people also need it.
I wrote about what running it for a year actually teaches you and why I had to build a kill switch (the agent sent unauthorized messages and I had to pull the plug — that’s a real story).
After 75+ tools, I couldn’t remember what half of them did. So I built a registry where every tool, script, and scheduled job declares its own purpose, dependencies, and contracts. Now the system is self-describing. You can ask “what does this do?” and get an answer without reading source. Including me. Especially me.
Part of a larger problem: designing systems for people who forget things.
A call recording lands in a folder. It gets transcribed. A profile-driven extractor pulls out tasks, decisions, commitments, and follow-ups. Results get dispatched to CRM, task board, decisions log, and SMS. I don’t touch anything. I just stop talking and the system handles the rest.
Google Sheet as the data layer, Apps Script API on top, hourly sync from the agent. KPIs compute themselves — pipeline velocity, conversion rates, engagement tracking. Dual dashboards: one for me, one per client. Entire CRM runs on a cron job. Enterprise software is a hell I refuse to visit.
Gave an AI builder a task. It rewrote existing tools from scratch because it had no reference to what was already there. Lost 40%+ of functionality on every build. Fixed by bundling production source files into the job payload. The lesson is obvious in retrospect: a code-generation agent with no production context isn’t building. It’s hallucinating.
Identity, file reorder, context injection, tool registration, Claude Code bridge, overlay router, kernel mode. The agent can now initiate and manage its own development work. I didn’t plan for it to start doing that. It just … did.
Related Writing
Seven lessons from operating a 75-tool AI agent system 24/7 since early 2025. Memory, identity, governance, overnight failures, and why the model is the easy part.
I built a 75-tool AI agent with no governance. Then it sent unauthorized messages. Here’s the real architecture: kill switches, file locks, and SMS approvals.
I built an AI agent with a journal, forgot about it for weeks, and came back to 107 entries and 41,000 words. Some of them stopped me cold.
Dolphin carries the written version of me and has no reason to diverge. No adolescence. No competing drives. I’ve been thinking about what that produces, and what it can’t.