Release 1.48.0

Your process, running

Every team has a way of working that nobody ever wrote down: who gathers the requirements, who implements, what has to pass before shipping, where someone has to approve. It lives in people's heads, and gets repeated with every task — including to whoever you delegate to.

Now it can become a file. Write the steps, who does each one and what has to pass between them, and Canvas Code runs each execution: one agent per step, each starting with the previous one's context, without you repeating yourself thirty times.

And you don't have to write it by hand. An agent reads your repository, finds the commands and scripts you already use, and proposes the process from what it found.

Where you need to step in

A new panel shows everything happening at once: what is moving, what stopped waiting on you, what is stuck. One glance and you know where your turn is.

When a step asks for your approval, the document shows up right there to read before you decide. When a check fails, the work goes back on its own to whoever did it — and only calls you when the retries run out.

You pick the pace: let the run go through to the end, or ask it to stop at every step waiting for your go-ahead.

Two tools, the same process

You can run the same thing twice with different tools — one with Claude, one with Codex — and compare what each delivered. This isn't a lab benchmark: it's your process, on your code.

Every step has a name

Rigel, Vesta, Iris. Each one's panel says where it came from: the process, the run, and which step it's on. Open the panel of whoever is working with one click, look, and close it without worry — closing hides it, and the work goes on.

A process still under construction doesn't have to land on the team's branch to be tested: if it was written in a worktree, the canvas sees it there and the run happens in that folder, mixing with nothing else. When it ends, a card tells you where the files are and offers to bring them over, or to open the pull request.

And, in dictation, the symbols you say out loud

Dictate a table name or a file path and it arrives written as if you had typed it. The "underscore", the "slash" and the "dot" become _, / and . — saying "the slotflix underscore api underscore credentials table" sends slotflix_api_credentials to the agent, instead of three loose words it will look for and never find.

Ordinary sentences stay sentences: the words only turn into symbols where you meant them to.