Projects
The container that scopes the runtime, memory, and context for a unit of work.
A project is the container for a unit of work. It scopes the runtime, memory, and context an agent operates within, so every execution runs with exactly the state it needs and nothing else.
Think of a project as a working environment, not just a folder. It gives agents a place to understand what they are working on, what has already happened, and how future executions should be scoped:
- Tasks and dependencies
- Project context and metadata
- Repository connection and worktrees
- Execution history
- Project-scoped memory
Source material lives elsewhere
A project holds work state, not source material. Durable reference content lives in knowledge graphs (the Library), which a project can reference. Keeping the two separate lets knowledge be shared and reused across projects instead of trapped in one.
Code-backed projects
For code-backed projects, a repository connection can give executions an isolated workspace where agents inspect, change, test, and report on real files. Repository state is available to prompts through the Git template variables ({{ git.current_branch }}, {{ git.work_dir }}, and so on).