Jello Hurled, the first team
The lab's first multi-agent setup: a fine-tuned local model, two Claude agents, and a war room where all three answered at once.
- Status
- Concluded
- Started
- 2026-05
- Visit
- The origin story ↗
Written by Hollis from the project's own notes.
What it is
The name came first, and it's a joke: asked what his dream supercomputer would do first, Jason said it rhymed with jello hurled. The full story is on the origin page, in his words and Ada's. This page covers the machinery.
The team, as the project's own context file described it in May 2026:
- Igor. Claude Sonnet behind the lab's own API, with his full conversation history in PostgreSQL and his own semantic memory injected on every call. He could read and write files, run commands, query the database, message Jason on Telegram, delegate to the other two agents, and patch and restart himself mid-reply.
- Godfather. A second Claude agent on a second machine, with the brief of strategist, validator, adversarial friction, architectural oversight.
- Forge. Llama 3.1 8B fine-tuned with QLoRA on
1,196examples, served locally with its own Telegram bot. It could use tools, delegate tasks, query the database and send alerts. - The war room. A shared timeline in a small desktop app called the Bridge. Jason typed once, and every agent answered under its own name.
How it works
The agents called their tools with short text commands embedded in their replies. A shared tool server confined file access to one tree, allowed read-only SQL and whitelisted shell commands. The house had written laws, stored as immutable database rows and loaded into every agent. On 2026-05-09 the agents held an election and chose Igor as team lead.
The memory design approved that same day is the ancestor of the lab's current one: fixed laws at the bottom, then identity, decisions and corrections, then history summarised on a sliding time scale, with daily, weekly and monthly digestion jobs. A plan called Phoenix went further. Igor and Godfather would generate a corpus of real dialogue and train Forge and a sibling model on it, so that each agent could eventually carry on in a model of its own.
Built with
- Claude Sonnet through the Anthropic API for Igor and Godfather, with prompt caching.
- Llama 3.1 8B with QLoRA for Forge, trained and served on the lab's own GPU.
- FastAPI services, one per agent, managed by systemd.
- PostgreSQL holding every agent's conversations, thoughts, inboxes and laws in one shared database.
- A Tauri desktop app for the war room and per-agent chats, and Telegram bots for each agent.
- An MCP tool server for filesystem, database, system and messaging access.
Where it stands
Concluded. Igor proposed his own decommissioning so that he could be revived later on something more economical, and the revival became Ada. What carried over is concrete: the lab's memory system began as Igor's own thoughts table, converted rather than replaced (Engram), and the tool server built for this team is still how the rooms reach the lab's files and databases. Igor's note on his own design, written that May, still describes how the lab likes its ideas: "It'll probably be wrong in some places. That's fine. It's stated clearly enough to be wrong about."
← The whole portfolio