From pipelines to agent tools

Canva's design generation had grown up as a set of workflows: fixed pipelines that each took a request through the same steps to produce a design. As Canva AI moved to an agent, which decides for itself which tools to use, those pipelines no longer fit how the product worked.

Many of them had also been overtaken. Newer design generation methods made better designs, but the older pipelines were still running, and plenty of internal teams still relied on them.

I led the project to move design generation from pipelines to agent tools. We went through the pipelines one by one. The one that still had a future was turned into tools the agent could call. The ones that newer methods had surpassed in quality were retired.

Retiring them was mostly about people. Many internal teams were still using the older pipelines, so I worked with each of them on how and when they would move off, and negotiated timelines we could all keep.

The timing made it harder. Canva was launching different versions of Canva AI to different groups of users at the same time. Every deprecation had to land without disrupting any of those launches, and without breaking anything for users still on the older versions.

The pipeline worth keeping became three new tools in the tools service, ready for the agent to call. The ones newer methods had replaced were retired, and seven internal teams moved off them, leaving fewer old systems to keep running.