agents / 02

Some progress on my personal agents.

Data

I believe self-hosted open weight models are best for everyday agents, since the durable thing is the structure that holds data and code (harness).

I bought a small Raspberry Pi to make this durable layer physical. It keeps a local Supabase instance, plus a synced copy of the repo. I can SSH into it from my computer, browse the data in Supabase Studio, and sync a read-only copy of the cloud database.

Following Timi’s setup, I also installed Tailscale on both computers. They can connect securely without being on the same Wi-Fi network.

A flash drive attached to my computer can probably a similar job, but I chose the Pi because I can extend it. For example, I’m going to design and 3D-print a case.

Goals

Captain now has a short goals file that defines what the agent should remember, study and do.

In summary: keep useful memory, stay close to my thinking, act through trusted tools and run careful experiments.

The current jobs are:

Learning

Curiosity study reads my essays and RSS feeds to discover new interesting ideas.
Rolodex expand scans my sources every week and adds new feeds that match those interests.
Liveness chats with me randomly to develop character.

Self-audit

Biodata sync reads new website notes every day and updates my biodata if needed.
Skills audit looks for newly relevant agent skills and emails the good matches to me.
Token usage review review daily token usage across agents and loops.
Daily review journals the previous day, refines my profile and interests, and sends a morning note when there is something useful to say.

Finance

Trading runs cautious Solana trading experiments within set risk limits.
Trading outcome review trading hypothesis and learn from result.