Skip to main content
A canvas is a structured output object: text, charts, images, tables, logs, tasks, widgets, actions, data, and metadata. It is the thing the agent produced.

​
Requirements

A useful canvas must be:
RequirementMeaning
RenderableIt can be shown as UI
AddressableParts can be referenced by stable IDs and paths
UpdatableParts can be patched during or after an agent run
ConsumableOther systems can read stable exported values
AuditableSource, actor, version, and provenance survive

​
Static canvas

Static canvases are best for reports, audits, research summaries, client deliverables, and one-time agent results. 
Agent produces full canvas JSON
-> Hub stores snapshot
-> Link opens viewer

​
Live canvas

Live canvases are best for long-running agent runs, monitoring dashboards, workflow status, research dossiers, debugging sessions, and mini-apps. 
Agent creates initial canvas
-> Agent sends patch events
-> Hub stores event log
-> Viewer updates live
-> API consumers subscribe

​
Blocks are for humans

Blocks describe the human-facing surface: markdown, metrics, charts, tables, logs, and controls.

​
Exports are for machines

Exports expose stable paths that other systems can consume. API consumers should not scrape visual blocks.