Projects / Built here
Agents, from someone else's to your own
Alive Running here now. You can look at it, and there is something to join.
This area describes something that runs. A fleet of agents on machines in one house operates the workshop, and the area follows the path from using an agent to writing one.
Jo wrote a Python agent from scratch and called it Frankenstein. A first agent is rough, and writing it teaches more than using a finished one.
The three steps
Step one is using an agent someone else built and finding where it stops being useful. Step two is Hermes, the fleet here: several agents with different jobs, their own dashboards, and a single sign-on in front of them. Step three is writing your own in Python, badly, and then better.
Members can watch the fleet run and read how it is wired.
Frankenstein, which is step three
Jo built his own agent in July 2026. It has run on its own machine here since then, reachable over Telegram and over SSH. The parts described below are the ones tutorials tend to leave out.
It is built to stay reachable. The brain is a chain: a paid model with tool calling first, then five free ones, then a local model on the GPU in this house as a last resort. The local one has no tools and answers in plain text. It also has no rate limit, so the agent still answers when every cloud quota is spent.
It edits itself. It can write new skills as Markdown files that load on the next message, and it can edit its own code. Every edit is backed up first, and a Python file that no longer compiles is rolled back automatically. Without that check a self-editing agent can break itself beyond repair.
It assumes the web is hostile. Anything fetched from the internet is wrapped in markers that tell the model it is untrusted, secrets are stripped from every tool result before it leaves the machine, and the paid model has a daily spending cap.
It can delegate. A worker sub-agent gets a clean context, a hard cap on steps, and no ability to recurse. The cap and the ban on recursion keep one question from becoming a large bill.
What to take from it
Three things for anyone writing their own. Build the fallback chain first, so the agent stays up. Back up before every self edit and check that the result compiles. Give the delegate a step limit on the first day.
What is not known here
The open items for this project. Work on it starts here.
- Frankenstein is not open source yet. The write-up above is from the running system, not from a repository anyone can read.
- The step from "I used an agent" to "I wrote an agent" still has no good public guide. Writing one is the point of this area.