Your Design Stack Is Growing Agents. Decide Where They Fit Before They Decide For You.

The tools your team uses to go from idea to shipped screen are folding coding agents right into the canvas. Figma bought a vibe-coding team. Claude's agent now runs on your phone. And there's a way to make agents build with your real components instead of made-up ones. Let me catch you up on what actually moved and what it means for your handoff flow.

The canvas is eating the code layer

Figma stopped acting like a place to draw pictures of apps. It acquired the team behind the vibe-coding platform Bud, a tool that spun up working apps for web, mobile, and Slack. Both Bud and Orchids shut down by July 18, so users had to migrate fast.

This follows a clear pattern. Figma shipped Figma Make for web apps, wired in Codex and Claude Code, and rolled out its own agents. The pitch is that ideas should not stop at a static mockup. They should turn into something that runs, inside the same tool your team already opens every day.

For you, that means the line between "design file" and "working prototype" is getting thin. Your designers will be able to build, not just draw. Decide now whether that changes who owns the first buildable version of a feature.

The agent left the coding lane

Coding was the opening act. The bigger story is agents moving into normal office work. Anthropic brought Claude Cowork to web and mobile, so you can start a task at your desk, check it on your phone, and pick up the output with your laptop closed.

The usage data tells the real story. Anthropic sampled 1.2 million Cowork sessions from more than 600,000 organizations. Software development was only 8.7% of use. Business operations was 33.4%, and content and copywriting was 16.4%. The work agents do best is the "work around the work," the reports and trackers and slide decks nobody's core job.

So the agents your team touches this quarter are not just for engineers. They are for your PM building a briefing doc and your designer turning call transcripts into a deck. Plan for that, not for a coding demo.

Made-up buttons versus your buttons

Here is the catch nobody puts on the launch slide. When an agent builds UI, it works from training data and your prompt. It has no idea what components your team already built. So it invents APIs and ignores your design system, and the output looks close enough at first glance while quietly ignoring decisions your codebase already made.

The fix is real and shipping. Storybook 10.3 added an MCP server that hands the agent your component library as tools it can query at runtime. In one test, the same prompt without it produced 263 lines of hand-rolled inputs. With it, the agent made six lookups first, then wrote 188 lines that imported the actual Button and TextInput, with every prop matching the documented contract.

It even caught its own accessibility bug, fixed the shared TextInput component, and re-ran the tests until all five passed. That fix helped every other screen using that component too.

The plumbing you don't see

Underneath the shiny tools, a fight is happening over how the pieces connect. Vercel's Guillermo Rauch says last year clients picked one lab and built everything on it. Now they treat model, harness, data platform, sandbox, and gateway as plug and play, swapping in Gemini or open models when the price and performance beat the headline name.

Rauch's other warning matters for your data. Point a coding agent at the wrong setting and your codebase can leave for training. He mentioned the president of Airbus and decades of aerospace C++ walking out the door because someone installed the wrong tool. Vercel's answer is a sandbox that caps what data an agent can touch and what can leave.

The deep cut

The agents are coming into your stack this quarter whether you plan for them or not. The one move that separates a mess from a win is grounding. An agent with no access to your component library builds plausible junk that a reviewer has to catch and reject. An agent wired to your design system, through something like the Storybook MCP server, builds with your real props and fixes its own bugs before you look.

So before you let anyone vibe-code a feature into your product, make your component library machine-readable and connect it. That one setup step turns an agent from a fast intern who ignores your standards into one who reads the docs first. Same tool, opposite outcome.

Three questions for your team

1. Who owns the first buildable version of a feature now that designers can generate working apps inside Figma? Draw that line before two people both think it's theirs.
2. Is our design system exposed to agents as queryable tools, or are we hoping they guess our components right? If it's the second one, that's this month's project.
3. What data can our coding agents actually reach, and can we audit it? Get a clear answer before an agent touches anything sensitive, not after.