AI Factory

• AVAILABLEAgent sheets and their lifecycleCreated in the app or drafted from the Terminal; approval is a click in the app.
• AVAILABLEPublishing a sheet for its roleA sheet in Production becomes selectable for its role, for the cards placed afterwards.
• PLANNEDTraining or fine-tuning a modelNot implemented. No training exists in Deskworth.

#An agent sheet

1. Open AI Factory: an agents tab and a models tab, with “New agent sheet” and “New model sheet”.
2. Create an agent sheet: a name, the compiled role it specialises, a narrower mandate, the capabilities, tools and sources it keeps, a provider, and limits under the desk’s caps. From the Terminal, the command writes a draft only.
3. Move it forward: Draft, Testing, Pending approval, Approved, Production, Archived. Approved and Production need your click. A sheet waiting for you carries the lit ring.
4. Publish it: a sheet in Production becomes selectable for its role in Settings › Agents, and applies to the cards placed afterwards.

deskworth factory create fundamental

#What a sheet can and cannot do

• A sheet specialises a compiled role and can only narrow it: capabilities, tools or sources outside the role are refused, and so are limits above the caps. Tests keep that true.
• The approval click can only come from the screen: neither the Terminal nor MCP can move a sheet forward. A test keeps that true.
• On a card, the sheet filters the tools twice: in what the agent is told it has, and at the moment it calls one.

#A model sheet

A model sheet describes a model; it does not train one. It records a method, a command-line tool, an API, fine-tuning with a provider, an open-weight model run locally, or a model from scratch. For fine-tuning with a provider, launching the job is not available yet; the from-scratch sheet models, estimates and documents.

LimitNo sheet is in Production in the recorded missions today: the next one waits on three sheets and six clicks from the operator. The target, a fund that creates, adapts and serves its own model, is not in the application yet, and no training exists.