The week in agent-readable design: Nov 10–17, 2026
This launch watch covers November 10 to 17, 2026. An agent learns to design only through loops it can read, write, and check. The thesis for this issue:
“This week’s only in-window launches — Kombai’s 2.0.179–2.0.185 changelog run and Figma’s 3D transform and Motion releases — show agent-readable design tooling consolidating around machine-callable commands with up-front costs and shareable check outputs, and they mark the field’s real dividing line: Kombai documented a callable asset pipeline and a linkable check report, while Figma widened what agents can produce with no documented way for an agent to verify it.”
Only two surfaces changed inside the window. Kombai’s changelog gained 2.0.179 (November 11), 2.0.180, 2.0.182 and 2.0.183 (November 12), 2.0.184 (November 13) and 2.0.185 (November 16), per the Kombai changelog. Figma shipped the Motion updates on November 10 and 3D transform on November 12, per the Figma release notes. Every other tool below is an earlier launch whose current documented surface was re-checked this week; none of them is an in-window launch.
Kombai: six changelog entries, one callable pipeline
Kombai’s six entries are the heaviest agent-surface change this week. Pre-window context: 2.0.172 (November 5, 2026) says the agent reads “every UI component from your codebase, npm packages, and Storybook.”
2.0.180 (November 12): the CLI grows an asset pipeline
The entry says the Kombai CLI “can now generate AI images and video for you directly from the command line.” Three subcommands are documented: kombai asset generate, kombai asset wait, and kombai asset models. Costs are “listed up front,” finished files come back “as a ready-to-use link, not just a local file path,” and generation happens only on request: “Kombai won’t quietly generate one on its own, even when a design could use the art.” A patch corrects credit estimates for “several image and video models (seedream, veo, kling, minimax, and others)” that had been quoted below actual charges.
2.0.179 (November 11): taste reports go shareable
Taste reports “now have a shareable web page at /t/<run>, linked from the CLI.” The same entry changes kombai taste start: it “now generates AI images by default alongside stock images (pass --no-generate to turn that off).” Figma frames can also be pasted into chat “with no canvas or Figma connection required.”
2.0.182, 2.0.183, 2.0.184, 2.0.185: interaction and fixes
In 2.0.182 questions from “a hosted Claude Code or Codex session” appear as “Kombai’s own interactive form instead of plain text,” resuming after submission, and the Install plugin action for Codex returns under Settings → Kombai MCP Server. 2.0.183 reads “Minor improvements and bug fixes.” 2.0.184 adds Gemini 4 Hydrogen and fixes a canvas staying “busy” for up to 10 minutes “after Kombai’s CLI finished checking it and found nothing to fix” — a completion state any check-then-report loop depends on. 2.0.185 adds Qwen 4 Max and makes the hosted canvas’s guest Sign in dialog “just ask you to have your agent sign you in.”
- Read — Documented. Component read from “codebase, npm packages, and Storybook” (2.0.172, pre-window); the library docs are linked, not described (Source).
- Write — Documented.
kombai asset generate,kombai asset wait,kombai asset modelsreturn ready-to-use links with costs up front (Source). - Verification — Documented with a stated caveat.
/t/<run>reports and a CLI check that “found nothing to fix” exist, but no Taste thresholds or pass/fail semantics are published (Source). - Access dependency — Documented. CLI, Settings → Kombai MCP Server, Codex plugin, and hosted Claude Code or Codex sessions (Source).
Figma: 3D transform (Nov 12) and Motion springs (Nov 10)
Figma’s two in-window releases widen what an agent can produce. Neither documents a way for the agent to check the result.
3D transform: agent-composable scenes behind a seat gate
The November 12 release, “Design in 3D, natively in Figma, and ship directly to code,” says 3D transform lets designers “rotate objects along the x, y, and z axes using on-canvas handles and live previews,” and “Build 3D designs from scratch, or let the Figma agent compose them for you.” The effect is non-destructive: “layers remain editable, and text and fills, including shader fills, render with the transformed layer.” Entitlements are “All plans,” “Requires a Full seat,” “Requires edit access to the file” (Figma help). The handoff is to “copy the CSS from Dev Mode or send designs directly to your coding agent with the Figma MCP.”
The Figma MCP server documents extraction of “variables, components, and layout data,” a figma-spec skill for team specs, and writes to the canvas. Protocol: “MCP servers use the Model Context Protocol, a standard interface that lets AI agents interact with data sources.” The write caveat, verbatim: “This feature requires the remote Figma MCP server and a Full seat to write outside your drafts. It’s free during the beta period and will eventually become a usage-based paid feature.”
Motion: springs become design-system artifacts
The November 10 release, “New Figma Motion updates: physics based springs, style and Lottie export previews,” describes springs where you “Set mass, stiffness, and damping for iOS and Android,” with “Save the spring in a design system to reuse at scale.” Style previews let you “Watch each style play in the animations list and on your design on canvas before you apply it,” and Lottie export previews show “how your animation plays as a Lottie file before you export it” (Figma release notes). Previews before export are not automated checks.
- Read — Documented. MCP extraction of “variables, components, and layout data,” plus
figma-spec(Source). - Write — Documented with a stated caveat. Canvas writes need the remote server and “a Full seat to write outside your drafts,” free “during the beta period” (Source).
- Verification — Not established by these sources. Previews and a Dev Mode CSS copy, not a check on agent output (Source).
- Access dependency — Documented. Remote MCP server plus a Full seat; 3D transform needs a Full seat and edit access (Source).
The verification gap this week exposed
Kombai is the only in-window surface with a source-documented check artifact: the /t/<run> report and the CLI check whose completion state 2.0.184 fixed. Even that is thin — no Taste documentation, thresholds, or pass/fail semantics appear, so an agent cannot be shown what a passing result means.
Figma widened what an agent can write, but its documented verification surface is previews and a Dev Mode CSS handoff. A preview is not a check. Write reach grew in Figma, check output appeared in Kombai, and the rest produce or preview, so the learning loop still lacks its third step across most of the market.
Re-checked: Modeinspect
An earlier launch, re-checked this week. Its homepage says: “Capture any screen of your live product onto the canvas. Pixel for pixel, fully editable.” Its enterprise page describes a read surface that “reads your file layout, components, tokens, conventions, and existing logic, then writes within them,” and “DeepCode retrieval keeps output high-fidelity across enterprise repos” (Modeinspect; Modeinspect enterprise). “Canvas to code” sends a frame back so it “lands in your codebase as a scoped diff, in your design system.” Verification is “Test the product — Click through real components on live data, at every breakpoint. Not a simulation.”
The metrics — “7× Faster design QA”, “Scoped, type-safe diffs that produce 80-90% of the code on complex features” — are vendor marketing claims, not findings of this site. No MCP, CLI, or API appears in these pages.
- Read — Documented. Live-screen capture plus reading of file layout, components, tokens (Source).
- Write — Documented. Canvas-to-code scoped diffs and PRs (Source).
- Verification — Documented with a stated caveat. Click-through is a vendor claim with no independent method (Source).
- Access dependency — Not established by these sources. No MCP, CLI, API or install docs (Source).
Re-checked: PixelCrew
An earlier launch, re-checked this week. A brief goes in and a crew of named agents hands work along a sequence: the research brief passes to Marcus, the creative brief to Mira. Each stage lists an output — wireframes.html, ux-flow.json, nav-spec, copy.md, tokens, landing-page.html, design-system.json, docs — and QA produces qa-report.md and revisions (PixelCrew). “Average delivery: 25–45 minutes” is a vendor claim, and the QA stage is self-reported, not externally audited. The site labels its case material: “Demonstration projects produced with PixelCrew from written briefs. Business names and brands shown are illustrative and are not clients or endorsements.”
- Read — Documented with a stated caveat. Inputs are the brief and upstream agent files; no codebase or design-file ingest is documented (Source).
- Write — Documented. Staged output files ending in a landing page, design-system JSON and docs (Source).
- Verification — Documented with a stated caveat. A
qa-report.mdstage, self-reported, with no external tooling documented (Source). - Access dependency — Documented. The web app at app.pixelcrew.ai; no MCP, CLI or API documented (Source).
Re-checked: ABrush
An earlier launch, re-checked this week. A Photoshop plugin whose flow runs from a selection, to a mode (Txt2Img, Img2Img, Upscale, Edit, Actions), to a model “(Stable Diffusion, Flux, Gwen Image and others),” to a prompt, to a render. The result goes to “Insert into document” as a Smart Object “at the coordinates of the zone that was sent,” and “You can queue up to 4 requests” (ABrush; ABrush docs). Pro Controls list “ControlNet support,” “IP-Adapter/References,” “Generation history,” “Presets and preset sharing,” and “LoRA styles.”
Verification stops at preview: “Review the preview, then press Insert into document.” The plugin is bound to Photoshop. No MCP, CLI or API is documented.
- Read — Documented with a stated caveat. Photoshop selection and captured images; no code, token or design-system read (Source).
- Write — Documented. Smart Object insert at the sent zone’s coordinates (Source).
- Verification — Not established by these sources. The documented loop ends at a preview (Source).
- Access dependency — Documented. The Photoshop plugin panel, sign-in and token balance (Source).
Re-checked: Impeccable, via the Designeer index
An earlier listing, re-checked this week, and only through its directory record. The Designeer tool page gives the category “AI & Agents” and this blurb: “Design intelligence skill for AI agents.” and “Turn AI slop into interfaces you’re proud to ship.” The record adds that the skill helps agents “build from scratch, improve what you have, and stay true to your design system.” The vendor link, impeccable.style, was linked but not reviewed, so no vendor page is described here; Designeer describes itself as an index of “422 tools across 17 sections, plus 132 design engineers worth following” (Designeer about).
- Read — Documented with a stated caveat. “stay true to your design system” implies a design-system read; one blurb, no mechanism (Source).
- Write — Documented with a stated caveat. “build from scratch, improve what you have”; no artifact type named (Source).
- Verification — Not established by these sources. Nothing describes checking output (Source).
- Access dependency — Documented with a stated caveat. A skill for AI agents, with no install command, host list or pricing (Source).
Re-checked: Shotcandy
An earlier open-source launch, re-checked this week. Its homepage says “Free and open source. No upload.” and offers paste or drop input for screenshots and screen recordings (Shotcandy). The repository shows 144 commits, a tests folder, and version 0.2.0 dated Oct 1, 2026. The in-window item is a commit dated Nov 10, 2026, “Upgrade Next.js to 16.4.5” (15fa803) — dependency maintenance, not a launch or product change. No MCP, CLI or API is documented.
- Read — Documented with a stated caveat. Pasted or dropped images and recordings only (Source).
- Write — Documented with a stated caveat. Beautified output in a web UI; no programmatic write path (Source).
- Verification — Not established by these sources. A tests folder exists; no verification workflow is described (Source).
- Access dependency — Documented. Free, open source, nothing uploaded (Source).
Resurf: unverified this week
Resurf is unverified this week. The only reachable record is a third-party directory listing, which calls it “A Personal Context Library for Mac” and says users can “hand that context to AI through MCP and CLI.” That is the directory’s claim, not a current fact. The listing records “Launched Sep 13, 2026,” 279 upvotes and 22 comments, and “Listing last verified: Sep 14, 2026,” and admits “Pricing details for this tool haven’t been verified yet” (Neuralpaws directory). The vendor site returned HTTP 500 (antibot) and the Product Hunt page was bot-walled as of 2026-11-17.
- Read — Not established by these sources (unverified this week). The MCP and CLI claim was last verified Sep 14, 2026 (Source).
- Write — Not established by these sources (unverified this week). No agent-facing write surface is recorded (Source).
- Verification — Not established by these sources (unverified this week). No verification surface appears (Source).
- Access dependency — Not established by these sources (unverified this week). Vendor site and Product Hunt unreachable (Source).
The launch-watch table
“Documented” means the source states the surface; “Documented with a stated caveat” means it qualifies it; “Not established by these sources” means the material does not support the claim.
| Tool | Read | Write | Verification | Access dependency |
|---|---|---|---|---|
| Kombai | Documented — 2.0.172 component-library claim (pre-window) | Documented — kombai asset generate|wait|models |
Documented with a stated caveat — /t/<run> report; no Taste thresholds |
Documented — CLI, MCP Server settings, hosted sessions |
| Figma | Documented — MCP design context, FigJam, Make | Documented with a stated caveat — Full seat, beta caveat | Not established by these sources — previews and Dev Mode CSS only | Documented — remote MCP, Full seat |
| Modeinspect | Documented — live-screen capture, tokens | Documented — scoped diffs, PRs | Documented with a stated caveat — vendor click-through claim | Not established by these sources — no MCP/CLI/API |
| PixelCrew | Documented with a stated caveat — brief and upstream files only | Documented — staged output files | Documented with a stated caveat — self-reported QA | Documented — web app |
| ABrush | Documented with a stated caveat — Photoshop selection only | Documented — Smart Object insert | Not established by these sources — preview only | Documented — Photoshop plugin |
| Impeccable | Documented with a stated caveat — directory blurb | Documented with a stated caveat — no artifact type | Not established by these sources — no check described | Documented with a stated caveat — skill, no install details |
| Shotcandy | Documented with a stated caveat — paste/drop input | Documented with a stated caveat — web UI output | Not established by these sources — tests folder only | Documented — free, open source, local |
| Resurf | Not established by these sources — (unverified this week) | Not established by these sources — (unverified this week) | Not established by these sources — (unverified this week) | Not established by these sources — (unverified this week) |
FAQ
What actually shipped between November 10 and 17, 2026? Kombai’s 2.0.179 (November 11), 2.0.180, 2.0.182 and 2.0.183 (November 12), 2.0.184 (November 13) and 2.0.185 (November 16), per the Kombai changelog; Figma’s Motion updates (November 10) and 3D transform (November 12), per the Figma release notes. Everything else here is a re-check of an earlier launch.
What can the Kombai CLI do after this week’s updates?
It has kombai asset generate, kombai asset wait and kombai asset models, with costs listed up front and ready-to-use links returned. kombai taste start generates AI images by default unless --no-generate is passed, and taste reports are shareable at /t/<run> (Kombai changelog).
Can Figma’s agent verify the 3D scenes it composes? Not according to these sources. The documented outputs are previews and Dev Mode CSS copies; Code Connect keeps “generated code consistent with your codebase,” a consistency aid rather than a check (Figma MCP docs).
Which of these tools can an agent call directly? Kombai documents a CLI, MCP and skill routes for “Cursor, Claude Code, Gemini CLI and Antigravity,” plus Codex plugin support (Kombai changelog). Figma documents a remote MCP server, with writes requiring a Full seat (Figma help). Modeinspect, PixelCrew, ABrush and Shotcandy document none; Impeccable is described as a skill with no install details, and Resurf is unverified.
Is Resurf’s MCP and CLI handoff available? Unverified as of 2026-11-17. Only a third-party directory, last verified Sep 14, 2026, makes the claim (Neuralpaws directory). The vendor site returned HTTP 500 and Product Hunt was bot-walled.
The Bottom Line
Agents learn to design only through loops they can read, write, and check. This week only Kombai added a documented check artifact — the /t/<run> report and a CLI check whose completion state was corrected (Kombai changelog). Figma widened the writable surface, including agent-composed 3D scenes, but left verification to humans and previews (Figma release notes). The re-checks show most of the market still ships produce-or-preview.
Lock-in signals, documented only: Figma’s Full-seat gate and the beta-free-then-paid MCP write path; ABrush’s binding to Adobe Photoshop; Shotcandy as the open-source, local-only counterpoint. Watch items: re-verify Resurf, watch for Kombai Taste documentation, and watch whether Figma documents a verification surface.
How This Guide Was Built
This is a documentation-only watch. All sources were fetched on 2026-11-17. The post involved no hands-on testing, no installs, and no vendor briefings. Linked-only pages, including the Kombai Component Library docs and Figma help sub-pages, are cited as links only.
Limits: the Kombai changelog excerpt is truncated after 2.0.169, the Figma MCP page is truncated at “Write to canva,” and Resurf was unreachable, so its entry rests on one directory record. Vendor metrics are quoted verbatim and attributed to vendor pages; they are not findings of this site.
