The year agents stop being a demo
The 2026 AI pragmatism narrative makes sense on the web. Mobile is still a cycle behind — the demos look good because the feedback loops are still carrying the weight.
The year agents stop being a demo
The early-January consensus from TechCrunch and Axios reads about right: 2026 is the year agentic workflows move from demo reel to daily practice. Gartner's 40% enterprise-apps-with-agents projection, the 340% jump in job postings requiring AI coding tool experience, Microsoft reporting 30% of its code written by AI — the trajectory isn't subtle.
On web, this is real. I've watched teams ship React features through Cursor and Claude Code at a cadence that would have been embarrassing to claim a year ago. The dev server reloads. The DOM is inspectable. The agent closes the loop on its own output. The demo is the product.
Mobile hasn't made that jump yet. The demos look good for the same reason any demo looks good — the reviewer fills in the gaps. A slick two-minute clip of an agent generating a SwiftUI screen doesn't say anything about what happens on day four, when the same agent needs to reason about a backstack that's three levels deep, a lifecycle method that fires twice on config change, or a Compose recomposition loop that only shows up on a mid-range Pixel.
The pragmatism framing is useful because it gets the question right. The question is no longer "can an agent write code." It's "can an agent close the loop on what it wrote." On web, mostly yes. On mobile, not yet — not without scaffolding the engineer builds around the model.
That's what makes the year interesting to me. The models are capable enough. The IDE integrations are landing. What's still missing on mobile is the substrate the agent needs to verify its own work: structured build output, runtime telemetry that flows back into context, test harnesses that produce signal the agent can act on.
If 2026 is the year agents stop being a demo, then on mobile it's the year the scaffolding problem stops being optional. The teams that build that layer get the leverage. The teams waiting for the model to do it alone will be writing the same "agent generated plausible Compose code that didn't work at runtime" post-mortem in December.
More on the specific pieces of that scaffolding over the next few posts. This is where the engineering work is.