GoKawiilEarlier this week, Figma introduced a way for AI agents to design directly in the Figma canvas based on your prompts. From Figma’s own blogpost:
With Figma’s MCP server, agents can now write directly to your Figma files, extending the standards you’ve carefully established over time. Via the use_figma tool, Claude Code, Codex, and other MCP clients can generate and modify design assets that are linked to your design system
What’s notable here is that Figma already had MCP tools, but they were read-only. In practice, this meant your AI agents could read your Figma files but they couldn’t write to them. This update gives them write access - a small update on the surface. Or is it?
If I had to guess, Figma was betting on their own AI creation tools like Figma Make to be the primary interface for “vibe-designing” - designing through prompting. From personal experience as well as that of many of my peers in the industry, the results from Figma Make are a resounding “meh”. It’s good demo-ware but fails pretty quickly as you try to get closer to any specific vision you have. It’s a tool most designers tried after their design VP mandated that everyone use AI tools in their work. Most didn’t return.
In parallel with all this, we’re seeing another fundamental shift in product development - more often than not, it now starts in Claude Code / Codex / Antigravity or whatever AI agent one may use. It’s now faster to prototype 10 directions with Claude Code than to mock up one wireframe in Figma. And that’s truly dangerous for Figma, because Figma is the place where product development used to start and live until engineering hand-off.
Figma, along with a large swath of SaaS tools can see both of these things clearly - their home-grown AI agents are collecting dust in their UI while people are falling head over heels for Claude Code and using it for everything.
There are two interesting powers at play here:
Claude Code becoming the consumer preference among builders and hence the integration point for other tools. In Ben Thompson’s framework, they’re becoming an aggregator. Claude Code’s differentiation is in deeply integrating their proprietary agent harness with their own models. Figma, Canva, Slack, and everyone else have access to the same model Anthropic does (Opus 4.6 being the current state-of-the-art) - yet, they can’t seem to replicate the magic of Claude Code within their own tools.
The immediate reaction of every SaaS company after they got caught like a deer in the AI headlights was to plug an AI agent into their app. “Our proprietary data is the secret sauce that will make our AI agent special” they thought. But that thought is now running into three harsh realities:
The most valuable data in any SaaS tool is your company’s context, which is almost always exposed with APIs for any other tool to use. Any metadata around actual usage patterns of said data & your tool is unlikely to be valuable enough to make any AI agent meaningfully more useful. Any individual SaaS company’s context is mostly useless without broader organizational context. Context has network effects - it gets exponentially more valuable the more complete it is and becomes almost useless when it’s limited. Your Figma files / Slack chats / Amplitude dashboards / JIRA tickets alone don’t give AI agents a complete picture of your business. But taken altogether, they’re extremely valuable. The software development process is moving from a discrete model - clear steps with explicit handoff at each step - to a more fluid model with no gates, stages, or handoffs as the process gets increasingly absorbed by the AI itself. Most SaaS tools were either designed to simplify a specific step and/or to improve the handoff between them. So what happens when those steps collapse into a process that needs no handoff?
... continue reading