Canva and the Specialist Flywheel

The specialist that compounds fastest is the one that gets invoked most often — because each invocation makes the specialist better, which attracts more invocations, which produces more improvements. This is the Arco Flywheel applied at the agent platform level, and it is what Canva has been building systematically for two and a half years. The qualifying sentence for this Perspective is what Canva’s head of ecosystem, Anwar Haneef, said at SaaStr Deploy 2026: “Every time an external agent reaches out to us to create a design, our platform gets smarter. Every time our user uses our agent in our own platform, it gets smarter. External agents are feeding one side of our intelligence layer and our own agents are feeding the other side. The platform that has the most learnings is the platform that can create the next best design.” This confirms what we argued in the Arco Flywheel and The Context We Built — that specialist platforms compound faster than generalist ones, and that operational experience accumulated at scale is the mechanism.

What Canva built and what the data shows

Canva, with more than 250 million monthly users and 27 billion uses of its AI capabilities to date, made the architectural decision most enterprise software companies are still deferring: rather than building a general creative agent, they became the undisputed visual design layer — the specialist every other agent calls when it needs something visual. The mechanism was the Machine-Readable Interface: an MCP server that makes Canva’s design capabilities accessible across every AI assistant and agent platform through a single standard protocol. The result is 30–40% month-over-month growth in MCP-connected usage across ChatGPT, Microsoft Copilot, and Claude — sustained over months, driven by the structural reality that when any agent needs a visual, Canva is the specialist it calls.

Two and a half years of production deployment produced five operational lessons. First: treat the MCP setup as a living system — “early structural decisions created problems we had to unwind,” the Rebuild Tax at the integration layer. Second: MCP is a surface, not a substitute for API design discipline — “an underlying API that is poorly scoped or ambiguous doesn’t disappear when you build MCP on top.” Third: agent experience requires first-principles design, not retrofitted human UX — “retrofitting a human experience doesn’t necessarily always translate for agents.” Fourth: trust architecture compounds — “if an agent doesn’t work well with your service, they’ll shut you down and won’t come back.” Fifth: the agentic landscape rewards specialists — “the tools that do one thing exceptionally well are the ones that succeed.”

The structural arguments this confirms

Three Arco structural arguments are confirmed simultaneously.

First, the Arco Flywheel at the platform level. Each invocation improves Canva’s design intelligence. A better specialist attracts more invocations. The flywheel accelerates and becomes progressively harder to displace — because the established specialist’s learning advantage widens with every additional invocation cycle. “Once that flywheel is spinning, it gets harder and harder to displace, because the platform that has the most learnings is the platform that can create the next best design.”

Second, the specialist design argument. Canva’s core strategic decision — “to be the undisputed visual design layer, so that when an agent creates something visual, they call Canva” — is the same structural principle that makes specialised autonomous businesses outperform diversified incumbents. When Eny built their agentic sales solution with Snowflake, they did not build their own design engine — they called Canva. The specialist that does one thing exceptionally well is the one that gets called. The question Canva’s head of ecosystem asked the audience is the right one for every agent design decision: “What is the one thing your product does that no one else does as well? Because in this agentic stack, the tools that do one thing exceptionally well are the ones that will actually succeed.”

Third, the De-SaaS-ing argument at the integration layer. “If your value is locked behind a graphical user interface, it is impossible for agents to find you.” Canva — with 250 million monthly users and strong commercial incentive to maintain the UI layer — is investing in the headless layer anyway, because the agent economy routes its calls there. The UI Tax that autonomous businesses were designed to eliminate from their own cost structures is the same tax that incumbent SaaS companies must remove from their integration surfaces to participate in the agent economy as called-upon specialists rather than bypassed incumbents.

The flywheel that Canva describes is the same flywheel the Arco Flywheel memo described at the portfolio level: each build generates intelligence that makes the next build better. At the platform level, the mechanism is identical — each design generated for an external agent makes the platform’s design intelligence better, which makes the next design better, which makes the specialist harder to replace. The businesses that build this flywheel in the specialist domain they own will not just serve the agent economy. They will compound within it.