The Agent panel on the canvas can now take a whole task: it reads what you already have on the canvas, plans, generates step by step, then looks at the output and decides whether to try again.
Tasks keep running in the background. Once you see "queued," collapse the panel or close the tab and pick up the progress when you return. The run shows live stages — reading the canvas, planning, executing, checking, correcting — and you can stop it anytime; anything already generating finishes and is billed on actual output.
It states the spend before it starts. Auto mode gives you a ceiling: at most 8 media nodes per task and 500 credits estimated. Cross either limit and nothing gets submitted.
Three workflow modes. Fast, guided, and production — review each stage before approving it. Character sheets come as a 2×2 grid, a horizontal three-view, or a body master. Scripts can be written fresh, adapted, or revised from your draft, from a 15-second short to five minutes plus.
It remembers your project. Tag a reference image with what it locks — facial identity, proportions, wardrobe, hair, expression, 12 roles in all — and the next mention of that character pulls the confirmed reference to hold the face. The Project Memory panel in the sidebar lets you read and rewrite what it knows.
It can watch and listen. Video and audio go in directly before it gives notes. It picks a generation model to match the task, and ships with skill libraries for storyboarding and prompt writing.
⌘I / Ctrl+I opens it.
Canvas Agent is now open to everyone — open any canvas and it's there. No request needed.