Figma Config 2026 Pixel Audit

Grant: Welcome back to UX Autopsy. I'm Grant. Maya: And I'm Maya. Alright, big week. Like, this is a big one. Grant: Figma dropped their design agent into beta on May 20th. Free for now, credits kick in at general availability. Maya: Which, yeah, that's the part we're going to dig into. Grant: So today's full autopsy, we've got the agent itself, the left rail UX, what it can actually do, what it can't. Maya: Then we're running red flag or green flag on three new Figma Make features. It's the PR workflow, tool call approvals, plan mode. Grant: And the predict pricing. Maya, the numbers on this are—look, I ran the math. Maya: Oh, No. Grant: Yeah, there's a cost asymmetry baked in here that I don't think most designers of clock care. Maya: We're also closing on Check Designs, Figma's June fourth QA tool, and there's a mental model in there worth stealing whether you're on a paid plan or not. Grant: Config 2026 is June 23rd through the 25th Moscone Center, so all of this is already in users' hands before the conference even opens. Maya: Which means Config isn't really a reveal this year, it's more like a ratification. Grant: Right, that's an interesting shift for how Figma does product launches. Maya: Okay, so the agent, it lives in the left rail now. Grant, you want to set the scene? Grant: So, Figma drops the design agent on May 20th. Professional, organization, enterprise seats, and the first thing I notice is the framing. The Figma blog says no AI credits during beta. Credits kick in at general availability. Maya: Which means everyone trying it right now is on a free trial they didn't negotiate. Grant: Exactly. And that framing matters. This isn't here's a product, it's here's a product. Bill comes later-classic soft launch with a deferred cost structure. Maya: Like a gym membership in January, you're in great shape until February. Grant: The math on that will get uncomfortable fast. But okay, what did they actually ship? Maya: So Figma's own blog says the agent lives in the left rail, acts directly on layers, and you can at-mention specific components and variables. It's not a sidebar generator, it edits the actual file. Grant: That distinction is real. Two years ago the pitch was, describe a screen, get a JPEG, nobody would ship. This is different: it reads your tokens, your component library, your variants. Maya: Right. The Muzli write up put it well: you can run prompts in parallel while you keep editing—it's not blocking. Grant: So architecturally, Figma's argument is: we built a teammate, not a generator. That's a meaningful claim. Maya: Hmm, I mean, I get why they're saying that, but a teammate implies shared understanding. The agent reads your design system, great. Does it understand why your card component is structured the way it is? That's a different thing. Grant: That's a fair distinction, reading tokens versus understanding intent. Maya: And then there's Config, June 23rd through 25th, Moscone Center, San Francisco, over 50 sessions, more than 75 design leaders. There's 8,000 attendees expected per UX trends listing. Grant: Which is the interesting wrinkle here. Config is supposed to be the big reveal stage, but the agent already shipped. So the keynote isn't an announcement, it's a ratification. Maya: Demo polish versus real-world readiness, those are two very different things to audit. Grant: And beta users will have had it for five weeks by the time the keynote hits. That's enough time for the cracks to show. Maya: Or enough time for Figma to quietly patch the embarrassing parts before the live stream. Grant: Right. Nothing like a conference to motivate a hot fix. Maya: So the agent is in your file, touching your layers, reading your design system. The real question isn't whether that's impressive, it's whether the way it works actually respects. respects how designers think or just approximates it from the outside. Grant: And that's worth pulling apart, because those two outcomes feel similar in a demo and completely different on a Tuesday afternoon deadline. So the left rail placement, that's the first tell. Maya: Oh, totally. It's not a popover and not a modal. It's sitting right where your layers live. Grant: And that's a real architectural decision, not just a design aesthetic. The Figma blog says the agent has deep context on your components, tokens, standards, things a third-party tool accessing via MCP just can't see. It's working from inside the file, not guessing from outside. Maya: Right. And Muzli had a good breakdown of this. They called it inside. Inside-out versus outside-in: external agents get a snapshot; the native agent gets your actual file structure. Grant: Which sounds great on paper. My issue is what happens when it gets something wrong. Maya: Okay. Grant: Because the agent edits layers directly. There's no visible reasoning. No, here's what I'm about to do. It just does it. And Figma's own help docs say you can undo, but the tip is to do. is to duplicate the frame first so you can compare before undoing. Maya: Wait, that's the official tip? Duplicate before you undo? Grant: Yeah. So the failure recovery workflow is prompt it, watch it change your file, realize something's off, duplicate the result, then undo the original. That's three steps to catch one mistake. Maya: I mean, I get why that bothers you, but compare that to what came before. Animas's blog put it well, two years of AI tools generating pretty mockups with wrong components outside any design system, nothing you'd actually ship. Grant: Sure, the bar was really low. Maya: Like, on the floor. So an agent that actually reads your tokens and respects your component library is a structural improvement, not a minor one. Grant: I don't dispute that. What I'm pushing on is the interaction model. Figma calls it a true collaborator, their exact words. But a collaborator tells you what they're doing. This thing operates as a black box at the layer level. Maya: To be fair, it's a beta. Grant: Every black box is a beta until it's a product. Maya: Okay, fair. Grant: And the early user reports reinforce this. Animas review flagged something that caught my attention. Designers, burning through sessions with the agent, mostly on self-correction cycles. You prompt it, it gets something wrong. It tries to fix itself, that burns more compute, and the designer ends up reviewing output rather than creating anything. Maya: Which flips the whole value proposition. Grant: Exactly. Youre not in a creative loop, youre in a review loop. Maya: Yeah. Grant: And the prompt interface is still natural language. Youre describing intent in words, not in design language. language. There's no way to say apply this specific component state without knowing how to phrase it so the agent interprets it correctly. Maya: Okay, but that's also true of any junior collaborator, right? You have to learn how they think. Grant: I'd accept that if the cost of miscommunication wasn't your actual file. Maya: That's the part I actually don't have a good answer for. Grant: It's a real tension. The design system awareness is genuinely new. Two: The interaction transparency-what the agent is doing, why and where the undo boundary sits-that's still unresolved. Maya: And I think reasonable people land differently on which of those matters more right now. Grant: Which is probably where we should leave it. Speaking of how you pay for each of those agent decisions, that feeds directly into the quick fire segment, where we're scoring some of Make's new moves one by one. All right, shifting gears, quick-fire round. Red flag or green? I'll go first. Let's do it. Maya: MAKE can now pull live web context mid-build via URL fetch. You approve each tool call before anything enters your session. Figma's own release notes confirm it. Tool call approvals let you review before anything enters your session. Red flag or green? Grant: Yellow, if I'm being honest. The approval gate is the right call on trust. I'd never want an agent silently pulling external content, but consent dialogs on a design canvas feel borrowed from a developer IDE. That's a context mismatch. Maya: So the mechanism is sound, but the UX metaphor is wrong. Grant: Exactly. Red flag on the experience, green flag on the intent. Maya: Fair. Okay, next one. Make connects to your local production codebase via the Desktop app: Desktop app: Figma's blog confirmed the May 28th beta drop. Designers can commit, Grant: And Maya: create a branch, open a pull request without touching a terminal. Grant: your engineering lead just felt a chill they can't explain. Maya: Right? So red flag or green? Grant: Big green flag on closing the handoff gap. I've watched designers spec something, hand it to an engineer, and watch it drift in translation for years. If the PR goes through normal review. The code doesn't bypass anyone. Bai Dai Yoda wrote a good piece on this, noting that changes still go through engineering review, not around it. That's the safeguard. Maya: My flag's more amber. Designers now need Git fluency, code base access, an engineering team willing to review their work. That's a real scope expansion. Someone's going to get hired for this, and their title is going to confuse everyone. Grant: The senior designer. Also, please know what a branch is. Maya: Exactly. Okay, last one. Plan mode. Figma's release notes describe it as an opt-in mode that asks clarifying questions, drafts a plan you can edit, and then you approve before anything gets built. But it costs more credits than a standard prompt, and it shows you an estimate before you commit. Grant: That estimate detail is doing a lot of work there. Maya: Right? So is that transparent pricing, or is it friction that trains you not to... not to explore. Grant: And the math on credits is where this gets uncomfortable. Showing an estimate sounds like transparency, but if complex work costs meaningfully more and the number is non-deterministic, which Figma's own help docs acknowledge, you're basically asking designers to gamble on a budget before they know what they're building. Maya: That's a red flag for me. Exploration is when AI tools are in their keep if the meter is visibly running you self-censor. Grant: Agreed. And honestly, that credit question doesn't end here. The pricing structure underneath all three of these flags is worth its own conversation. Maya: Which is exactly where we're going. Grant: So the credit math, because I've been staring at the numbers, and they don't add up. Maya: Okay, lay it out. Grant: On the Professional plan, five thousand extra credits cost one hundred and fifty dollars a month, adding a full seat that's sixteen dollars, and that seat comes with three thousand credits. Someone on the Figma forum ran the actual arithmetic, buying official extra credits is five point six times more expensive per credit. Then spinning up a goat seat. Maya: Wait, so the cheapest move is to add a fake employee? Grant: That's the implied incentive. I've read deal terms that accidentally created perverse outcomes before, but usually the lawyers catch it. This one is in plain sight. Maya: I mean, come on, that's wild, right? Grant: And the consumption side is just as messy. Figma's own help docs say credit usage is non-deterministic, meaning even Figma can't tell you what a prompt costs before you run it. One complex make prompt anywhere from 60 to 340 credits. Same task, different day, different bill. Maya: So you literally cannot budget for exploratory work. Grant: Which is exactly the use case. Nobody burns 300 credits on a prompt they know will work. You burn them figuring out if the prompt can work. Maya: Hmm, okay, so, and I want to push back a little here. The beta was free! Figma gave people months of actual free access. Doesn't that buy them some goodwill on the pricing rollout? Grant: Somewhat. The free beta period is real. But it also trained designers to use it exploratively, and now the meter penalizes exactly that behavior. You can't separate those two things. Maya: I don't fully disagree. I just think, like, enterprise teams have procurement budgets for this stuff. For them, $150 a month is noise. Maybe the model works fine at that scale. Grant: Sure, and that's kind of the point I keep coming back to. Look at who benefits from this structure. The answer isn't the freelancer or professional. And it's not the solo designer iterating on a side project. It's the org with the procurement team that can negotiate a custom rate. The credit system is shaped for that customer. Maya: Which maybe isn't wrong, like enterprise paying for AI at scale while individuals get the free tier to experiment. Grant: Except the free tier is 500 credits a month. Figma's own numbers put 3000 credits at roughly 50 to 70 prompts. Scale that to 500 and you get, what, eight prompts before you hit the wall? Maya: Yeah, okay, that's not experimentation, that's a demo. Grant: And I'm not saying Figma is wrong to charge for AI compute, that cost is real. But the way the tiers are structured, AI reads less like a platform capability and more like a metered utility, closer to AWS than to a design tool. Maya: Whereas if they'd bundled it flat into the seat. Seat price and eat in the margin for a year? The story's completely different. Grant: The math probably doesn't, but the perception definitely does. Maya: Look, I don't think we're going to land on the same side here. Grant: That's fine. I think the Beta Goodwill bought them more patience than Grant is giving them. Grant thinks the structure tells you who the real customer is. Maya: I think both things are true. Grant: Fair enough. Anyway, shifting gears completely, there's one Config announcement that sidesteps this whole credit debate, and it's the one I actually want to steal from. Check Designs dropped June 4th, and we should talk about it. Maya: Sure. Grant: Okay, so, shifting gears, Check Designs. Figma dropped it June 4th and it's the quietest announcement from this whole Config cycle. Maya: And probably the most immediately usable. No credits, no beta waitlist, it just runs. Grant: Right. So what it does, according to the Figma forum post, flags hard-coded color, text, radius, and spacing values, swaps them for the correct direct design system token in one click also catches WCAG 2.0 AA and AAA contrast violations, detached components, library mismatches. Maya: That's a linting pass. That's literally what this is, pre-hand off linting for your design file. Grant: Yes, and that's the mental model I want every designer to steal, even if you can't touch Czech designs. Maya: One catch worth naming, though. It's organization and enterprise only, Which Grant: Hmm. Maya: is almost weirdly coherent. The team's drowning in token drift and multim-library chaos are exactly the ones on those plans. Grant: Okay, I'll give you that. The problem and the gate actually align here. Maya: Unlike the credit model which we just... yeah. Grant: We're not going back there. But the point stands, even if you're on a Professional plan or flying solo, the idea transfers. Before handoff, do a linting pass. Not a vibe check, not does it look right. Systematically ask, are there hard-coded values, Detached components, contrast violations? Maya: Treat it like code review for your file. Grant: Exactly. That reframe, QA as design system health check, not pixel matching. That's the one thing worth walking away from this whole config drop. Maya: Even if you're doing it manually with a plug-in or just a checklist. Grant: Especially then. Maya: All right, that's a wrap on the Figma design agent audit. Grant: And what a one. I keep coming back to that deferred cost framing Grant. Free beta now, bill at GA. It's clever. Maya: Right. And a thing that sticks with me, the agent reads your design system, sure, but does it understand why your car component is structured the way it is? Your question still unanswered. Grant: I'm putting it on a sticky note. Honestly, the takeaway here is... Is "treat that pre-handoff pass like a linting run, not a visual review." That one's portable regardless of what plan you're on. Maya: Exactly. Check designs made that concrete. Token drift, contrast violations, detached components, catch it before it ships. Grant: So the homework? Pull up one file this week and run that mental linting pass. Maya: Screenshot a UI that bugs you, tag us, and subscribe so you catch next week's audit. Grant: Leave a review if today made you side-eye your own layout. Layer panel. Maya: Configs in a week. June twenty third, Moscone Center. We'll be watching. Grant: See you there or, you know, on the Internet. Thanks for listening.