Agents and workspaces
AI agents build. Each feature gets its own workspace, and the model and effort follow the risk of the change.
What it is
An AI agent here is a model, an effort level and a set of tools. Methods uses a cheaper model for small changes, and a stronger model or a second attempt for risky ones.
Agents are not tied to one provider. Claude, OpenAI, local models, deterministic scripts and people can all fill the same role.
How it works
Pick a ready feature
The scheduler takes a feature whose dependencies have been merged.
Choose an agent
The model and effort follow the attempt number and the risk of the change, not a global default.
Open a workspace
The feature is built in an isolated workspace of its own, so features built in parallel cannot disturb each other.
Build within the allowed files
The agent plans, writes the code and runs fast local checks, touching only the files the contract allows. The result goes to verification, not straight to the repository.
What it guarantees
- One feature per workspace, so work in progress is isolated.
- The builder never grades its own work. Verification is a separate step.
- No particular agent is required. Any role can be filled by a different model, a script or a person.
What is not built yet
- No agent runs yet, so no model has built a feature through Methods.
- The rules that match a kind of change to a model are not written.
- Isolated workspaces are designed, not implemented.
Build with us as a design partner.
Methods is pre-alpha. A small number of teams will run it on real work while it is built, and shape what it becomes.