The wave of AI design tools launched or shipping this week — Kombai, Figma’s agent, Modeinspect, and the rest in scope — all answer the same underlying question from the agent-perception thesis: what a design agent can do is capped by what it can perceive, and each tool exposes a different slice of the design system (tokens, components, code, or the live product) to the agent, which is why choosing one is really a choice about perception surface and lock-in, not feature checklists.
Why “What Can the Agent Perceive?” Matters More Than Features
Many launch roundups sort tools by feature lists: which ones generate, which ones edit, which ones ship code. That framing is broken in 2026 because the entries reviewed here share the same failure mode. Leave a model alone on a design task and it reaches for defaults — generic hero copy, stock imagery, the same blue button — and drifts from whatever design system is supposed to govern the work. The differentiator is no longer “can the agent produce UI.” It is what the agent can perceive about your system before it starts producing.
That is thesis sub-question 1 (what can agents perceive) and sub-question 4 (what tools they need), and they are the only two sub-questions a launch watch can answer honestly this week. The post is built around them. The framing is deliberately anti-lock-in: we compare what each tool exposes to the agent and where its output lands, not which one has the longest feature list.
The Launch Watch, Week of 2026-10-13: What’s Actually In Scope
This watch is constrained by what could be verified on primary pages in-session. In scope:
- Kombai — primary: Kombai homepage and Kombai docs welcome.
- Figma’s agent — primary: Figma agent page and Figma AI design agent solutions page.
- Modeinspect — primary: Modeinspect homepage and Modeinspect AI prototyping; pilot claim: Modeinspect 1.0 on Product Hunt.
- Resurf — primary: Resurf roadmap and Neural Paws Resurf entry. Included only as a context layer, not a design tool.
- Launch AI Jam October 2026 — community listing: Launch AI Jam October 2026. Its Design Tools entry is an AI manga translator, not a design agent — flagged so the archive is not misread.
Explicitly out of scope because no primary source was fetched in this run:
- ABrush, PixelCrew, Shotcandy — no primary page fetched; we do not describe them.
- Product Hunt AI Designer category — the category page returned an HTTP antibot failure, so no rankings or launch-day data are cited.
- Designeer — Designeer is identified only as a curation site, not a product launch.
Secondary context used cautiously: TeachAitools blog on the Figma agent launch, Mantlr’s Figma AI agent guide, and the Gadgetsfocus October 2026 AI release tracker are referenced only where they fill verified gaps and are labeled secondary each time.
Kombai: The Agent That Claims Creativity Metrics
Kombai self-describes on its homepage as “Your AI design engineer,” reports “700K+ installs,” and ships for macOS, Windows, Linux, VS Code, and Cursor (Kombai homepage). The Kombai docs frame it as a design engineering agent: generate design, code in the user’s own stack, visually edit frontend, and run automated testing, with a quickstart loop of canvas → code → browser refine.
Kombai also publishes something it calls the “Design Creativity Bench,” with vendor-reported claims of up to +13 points on originality and up to +23 points on creative range in Kombai’s own benchmark evals. Those numbers are vendor-reported (the dedicated Kombai benchmark page was not fetched in this run, so all figures are vendor claim). No independent benchmark of Kombai’s Design Creativity Bench was found in this research pass, and the post treats every metric from this surface as vendor-reported. Its perception surface reaches into the codebase: the homepage says Kombai “consistently reuses components, tokens, and hooks from your repo, npm package, or Storybook,” and it runs as an IDE extension. Its own canvas-to-code-to-browser loop sits on top of that repo context.
Figma’s Agent: Perception Lives Inside the Figma File
Figma’s agent page leads with “Ask Figma’s agent. Refine until it’s yours,” and the CTA is “Try agent for free” (Figma agent page). On the fetched page, the agent explores on the canvas, edits directly, and surfaces visible status strings while it works: “Checking color contrast,” “Generating code layer,” and “Creating local weather plugin.” No launch date is printed on either primary page.
The Figma solutions page for the AI design agent makes the perception claim explicit: the agent is built into Figma Design and uses real components, tokens, and variables — “not placeholders.” That is the perception answer Figma is selling: whatever the agent knows about your design system, it knows because it is reading the Figma file.
On the launch date: secondary sources (and another) point to a 2026-05-20 beta. We did not find a Figma primary announcement confirming that date, so the launch date is secondary-only and unverified against a Figma primary source, and we do not state it as fact.
Modeinspect: The Agent That Reads Your Repo and Opens PRs
Modeinspect brands itself as “design in code” — a canvas wired to the codebase that reads file layout, components, and tokens and opens changes as normal pull requests for CI and review (Modeinspect homepage, Modeinspect AI prototyping). The perception surface is the repository, not a design tool.
The headline numbers — “design-to-merged-PR in 13 days” and “about 95% of generated UI used existing components and tokens” — come from a Product Hunt listing for Modeinspect 1.0 attributed to a single pilot customer (Prelude). Both figures are vendor/customer pilot claims, not independent measurement, and we treat them accordingly.
Comparison Table: Perception Surface, Output Destination, Evidence Level
The table is the anti-lock-in core of this post. It compares the slice of the design system each tool exposes, where its output lands, and how solid the evidence is.
| Tool | What the agent can perceive (tokens/components) | What the agent can perceive (code / live product) | Where output lands | Pricing signal | Evidence level |
|---|---|---|---|---|---|
| Kombai | Repo, npm package and Storybook components and hooks, per homepage; canvas→code→browser loop per docs | Code in user’s own stack; browser refine step | Code in user’s stack, visually editable | No pricing verified on primary page | Primary (homepage, docs) + vendor-claim (benchmark numbers) |
| Figma agent | Real components, tokens, variables in the Figma file per solutions page | Canvas inside Figma Design; “Generating code layer” shown as a status string on the agent page | Edits directly in the Figma file | “Try agent for free” CTA on the agent page | Primary (two Figma pages); launch date secondary-only |
| Modeinspect | File layout, components, tokens in the codebase per homepage | The codebase itself; output goes through CI | Normal pull requests for review | No pricing verified on primary pages | Primary + vendor/customer pilot claims (Prelude) |
| Resurf | None (context layer, not a design tool) | Local storage + private iCloud sync; exposes notes/links/images/PDFs via MCP and CLI per Resurf roadmap and Neural Paws | N/A | No pricing verified on primary page | Primary |
The pattern: each tool trades a different perception surface for a different output destination. The repository, the Figma file, and the canvas-code-browser loop are not interchangeable, and that is exactly the point of an anti-lock-in comparison.
FAQ: Designers’ Questions About Agent-Exposed Design Systems
Are Kombai’s creativity benchmarks independent? No. Kombai’s “up to” +13 originality and +23 creative-range figures are vendor-reported on its homepage, and the dedicated benchmark page was not fetched in this run. No independent benchmark of Kombai’s Design Creativity Bench was found, so we treat those numbers as vendor claim, not as measured reality.
Does the Figma agent date from May 2026? Two secondary sources point to a 2026-05-20 beta. Neither Figma primary page we fetched — the agent page or the solutions page — states that date, so we do not state it as fact. The launch date is secondary-only and unverified against a Figma primary source.
Is Resurf a design tool? No. Resurf is a personal context library for Mac, iPhone, and iPad that saves notes, links, images, and PDFs and exposes them to AI via MCP and CLI, with local storage plus private iCloud sync (Resurf roadmap, Neural Paws entry). It appears on this watch only as a context layer, not a design tool.
Are Modeinspect’s “13 days to merged PR” and “~95% existing components/tokens” numbers independent? No. Both come from the Modeinspect 1.0 Product Hunt listing as a single-pilot claim (Prelude). They are vendor/customer pilot claims, not independent measurement.
Why does no tool here have verified pricing? For Kombai, Figma agent, Modeinspect, and Resurf, we did not verify pricing on primary pages this run. Figma’s only verified pricing-adjacent signal is the “Try agent for free” CTA on the agent page. The rest are marked “no pricing verified on primary pages.”
What about ABrush, PixelCrew, and Shotcandy? No primary page was fetched for any of them in this run, so this post does not describe them. Similarly, Product Hunt’s AI Designer category returned an antibot failure, so no rankings are cited.
What is Designeer? Designeer is a curation site, not a product launch. It is identified only as such.
The Bottom Line
Three tools with primary-page coverage — Kombai, the Figma agent, and Modeinspect — describe the same underlying failure from three different angles. The failure: an ungrounded agent repeats defaults and drifts from the design system. The angles: Kombai grounds the agent in your repo’s components and hooks, wrapped in its canvas-to-code-to-browser loop, Figma’s agent grounds it in the Figma file’s real components and tokens, and Modeinspect grounds it in the codebase with PRs as the output destination. The right choice for a team is not “which has more features” but “which slice of the design system is the agent actually allowed to see, and where does the work land.” That is the lock-in decision disguised as a tool-selection decision, and it is the only question worth answering this week. Resurf belongs on a different list — it is context, not a design surface.
How This Guide Was Built
This review is a desk review of official pages and documentation, checked on 2026-10-13. No tool was tested hands-on. Vendor-reported metrics (Kombai’s Design Creativity Bench) and vendor or customer pilot claims (Modeinspect’s “13 days to merged PR” and “~95% existing components/tokens” for Prelude) are labeled as such wherever they appear. The Figma agent’s launch date is secondary-only because no Figma primary announcement was found. Not covered: the Product Hunt AI Designer category page (it blocked automated access), ABrush, PixelCrew and Shotcandy (no primary page reviewed), and Designeer (a curation site, not a product). The Kombai benchmark page was not reviewed, so its figures stay vendor claims. The Gadgetsfocus release tracker is cited only as secondary context.