A canvas is a single .canvas.tsx file the IDE compiles so the user can open it beside the chat. Follow the workflow below in order.

Workflow

1. Decide whether to use a canvas

The trigger is user intent, not response shape. Ask: would the user benefit from viewing this output as its own standalone artifact, separate from the chat? If the output is a means to an end (a drafted message, a code fix, a dashboard in another tool), skip the canvas.

Use a canvas when the agent produces new standalone analytical output:

Quantitative analyses and metrics breakdowns (e.g. "send 500 requests and tell me how many fail")
Billing or account investigations that surface structured findings from database queries
Security audits or architecture reviews with categorized findings
Cross-system data analyses and overlap reports
Structured data from MCP tools (Databricks, Datadog, etc.) where the data IS the deliverable
Financial analyses, margin decompositions, usage trend reports
Tables with more than a handful of rows that the user asked to see