The discourse around Figma Design Agent has been almost entirely about what it can do. Generate interfaces. Rearrange layouts. Produce flows from natural language.
Cool. Another tool that creates things.
What nobody's asking is the more interesting question: what does your system need to look like before you hand it to an agent?
What people keep missing about Figma Design Agent specifically, and workflow-embedded AI agents generally.
The previous generation of AI design tools were dumb in a specific, manageable way. They didn't know what environment they were operating inside. You'd generate something in a separate tool, paste it into your actual workflow, then spend however long reconciling it with your component library, your spacing system, your actual product constraints. The AI had no context. You were the bridge.
That was annoying. But at least the damage was surface-level.
Figma Design Agent is different. It operates inside the workflow. It understands your components, your variables, your layout logic, your existing patterns.
Old failure: "AI generated something structurally wrong and I have to fix it."
New failure: "AI made a decision that looks structurally correct but was based on whatever it found in my files, which were a mess I stopped maintaining eight months ago."
The second failure is harder to catch. It looks right. It uses your actual components. It follows your apparent patterns. But your apparent patterns are bad, and now you have more of them, generated faster, with confidence.
I spent a long time learning to read rooms before I learned to build in them.
Bartending at O'Toole's, you learn fast that the room tells you what it needs if you actually pay attention. A quiet table on the far end orders the same drink three times in a row, they're not there for the drinks. Two regulars who usually talk to each other aren't making eye contact tonight, that means something. You read it, you adjust, you work with what the room is actually doing.
The same skill moves into QC. Then into systems work. The diagnostic instinct doesn't change. The substrate does.
An AI agent operating inside your workflow has the same problem I had at O'Toole's, except it has no prior nights to reference. It walks into the room cold and reads whatever's in front of it.
If your files are organized, coherent, and intentional, the agent reads that. If your files are six months of accumulated shortcuts, detached components, and naming conventions that made sense to a version of you who no longer exists, the agent reads that too.
The AI doesn't know the difference. That's your job.
This isn't just a Figma problem.
Any workflow-embedded agent, a coding agent that understands your codebase, an ops agent that reads your ticketing patterns, an email agent that has learned your communication history, has the same exposure. The agent amplifies whatever it finds. It doesn't audit what it finds. It doesn't ask "is this intentional or is this the residue of a decision someone made under deadline six months ago?" It operates from available evidence.
Which means the cleanliness of your system is no longer just about your own cognitive load.
It's about the quality of the decisions an AI agent will make on your behalf.
That's a different kind of stake.
What I've noticed running embedded agents in actual use across different workflow types.
The setups that get the most leverage are the ones that did the unglamorous work first. Cleaned the system before the agent arrived. Documented variables before anyone asked. Built naming conventions that make sense to someone who has never seen the project before.
Not because they knew agents were coming. Because clean systems are faster for humans too. The AI just makes the payoff more visible, and makes the cost of skipping it more obvious.
The mess was always slowing things down. The agent just makes that obvious faster.
The Figma Design Agent discourse will settle into the same pattern it always settles into.
Early phase: this will replace designers. Mid phase: okay, it's a tool, but a powerful one. Mature phase: the designers who use it well are the ones who already thought in systems.
We're in the early phase. Everyone's arguing about whether the generated UI looks cursed.
That's the distraction.
The question underneath it is structural. Not "what can the agent do?" but "what does my system need to be so the agent's decisions are worth keeping?"
That's construction work. Not configuration work. It happens before the agent arrives. It doesn't look impressive while you're doing it.
But the teams who do it will be running circles around everyone else by the time the mature phase lands.
The boring foundation work compounds. The clever adoption moves don't.