Assets

An asset is a thing you would miss. The test is loss: if it leaking, breaking, or changing behind your back would matter, it's an asset. Your Gmail account is an asset. The invoice record, the note the agent maintains, the program your team reuses, the view component three scenes share — assets. They have owners, permissions, approval gates, audit history, and a place in blueprint exports.

An asset has a type, fields, and an identity

• A type — a string like github/repository, gmail/account, daslab/note. The type determines what the asset can do: its tools, fields, views.
• A parent chain — a PR belongs to a repo belongs to a GitHub account. Most assets resolve their credentials by walking up this chain.
• Fields — typed, schema-defined properties (the repo's name, default branch, last commit SHA).
• An external_id — the canonical ID in the source system, when there is one.

The asset is the single source of truth. Ten cells across five scenes can project it; the fields live in one place, and every projection follows.

Two scopes: world and scene

A world is the root scene — the top of your workspace, with every other scene nested under it. Asset types declare their scope against that ladder (on the wire, literally scope: "world"):

• world-level — accounts, credentials, things owned by the world and shared across every scene under it. Connect Gmail once; every scene that needs it uses the same connection.
• scene-level — content that lives in one specific scene. A note, a record, a tracked filter. The default when not declared.

The split determines who sees what. World-level assets are visible to anyone in the world with the right role; scene-level assets only to people with access to that scene.

Tools and gates

An asset's type declares the tools that operate on it — github_list_pulls, gmail.send. Reads run freely. Writes that touch the real world stop at an approval gate. The governed surface is exactly the asset surface — rearranging a scene never asks permission, and sending an email always does.

Assets arrive by browse, by tool, or by hook

Pick one from a connected account's catalog, let the agent create one with a tool call, or let a lifecycle event (a job completing, a schedule firing) produce one as a side effect.

Removing a cell doesn't delete the asset

An asset referenced from many scenes has one set of fields. Remove its cell from one scene and you've stopped looking at it there — the asset, and every other scene's view of it, is untouched. Deleting the asset itself is the governed act.

What's next

• Cells — how scenes show assets.
• Providers — where asset types come from.
• Approvals — the gates on asset writes.
• Blueprints — assets declared as data, credentials sealed.

Updated 2026-07-18