Agent capability map

How North Star discovers tools, context, sources, and execution boundaries

North Star may only promise a capability after the current turn admits a real registered tool, source, permission, binding, credit policy, and durable evidence path. A feature visible in navigation or a repository Markdown file is not automatically agent context.

Source resolution order

For Taskcenter behavior, North Star searches and reads the bundled product docs first. For repeatable procedure it searches and reads skills. For workspace facts it uses Vectorize-backed semantic memory only over records that have actually been indexed and pass access checks. For a supplied public URL it uses URL extraction; for open-web discovery it uses live web search; deep research needs a confirmed brief and explicit approval. This order prevents public web search from replacing a document, attachment, or project record the user already supplied.

What the agent can and cannot know automatically

Authored/generated Taskcenter docs are directly searchable through searchDocs/readDocs. Workspace semantic memory is real but partial: it contains explicitly indexed records such as approved storyboard plans and issues, not every R2 file, upload, project document, Git repository Markdown file, or arbitrary internet page. A user should attach or link the material they want considered; the agent must state when it cannot retrieve or inspect it. The current general-loop source resolver is still under repair, so tool registration alone does not guarantee that an arbitrary chat request will select the correct internal source.

Capability status

Quick public web search through Parallel is live and production-verified for the native Agent path. URL extraction, deep research, connectors, routines, project operations, browser/sandbox work, decks, documents, media, spreadsheets, and websites each have registered or isolated execution paths, but their availability remains turn-, credential-, permission-, credit-, approval-, and evidence-dependent. Project CEO and child-agent orchestration are project-site-only, never a hidden general-chat action.

Adding an agent-operable feature

A new feature needs a shared ToolDefinition and toolset, capability metadata with an explicit permission, a guarded D1/R2 or provider path, pre-spend credits and approvals where needed, a source/context policy, durable events and artifacts, an honest empty/failure state, a product docs page, and production evidence. UI shortcuts that only write prompt text do not make a feature executable.

Current foundation limit

The general Agent foundation is still under release hold. The runtime currently uses deterministic capability policies alongside the model loop, and broad capability inventories are not yet the final bounded model tool surface. North Star must not claim a capability is complete merely because it appears in this map; the current admitted turn and its retained evidence are the source of truth.