Tencent's Miora Design Agent Opens to All: Design Agents Enter the Harness Stage

> Source: https://mp.weixin.qq.com/s/qQhq9nxoeCD68iMwQoEVFQ > Score: 77 · Category: ai-products · Published: 2026-07-22

Tencent's design agent platform Miora became fully open on July 22. Previously it required an invitation code and a waitlist; now it's open to everyone. The real significance isn't "another design AI launched"—it's that Miora pushes the "design agent" category from "one-prompt image generation" toward "complete brand engineering."

At miora.design, the interface shows four columns on the left (Create / Inspiration / Skills / Memory), a conversation area in the middle, and controls in the lower right for switching the agent base model (Standard / Pro / Max inference tiers) and multimodal generation models (image, video, UI, and 3D each selected independently). This is Miora's core structure.

Three Things That Set It Apart

1. Agent base model and multimodal models are separated.

Most design agents (Canva AI, Adobe Firefly ecosystem, and domestic equivalents) default to "one all-purpose model + fixed downstream outputs." Miora instead lets the agent handle reasoning and decisions (choosing Pro or Max), while image / video / UI / 3D tasks each pick their own model. NanoBanana, GPT-image-2, Seedance 2.0, Kling, and Vidu Q3 can all be invoked in the same conversation.

This "controller + pluggable multimodal" design turns model composition into a product-level option. The cost is that users must understand model differences; the benefit is user-directed multi-model collaboration instead of "one uniform AI-generated flavor."

2. A Skill marketplace.

Miora shares its Skill architecture with WorkBuddy. Every "template" (brand visual identity packages, logo generators, interactive game assets) is essentially a Skill. Users can create their own Skills—for example, selecting a brand scheme, generating a "brand VI Skill," and publishing it to the marketplace.

This mirrors the product form of Claude Code Skills and Anthropic's Skill SDK: prompts and engineering settings become reusable, packaged workflows rather than one-off inputs.

3. Layered memory.

Miora's memory system splits into project memory and personal memory, similar to Claude Code's project + user directory structure. Project memory covers progress / rules / assets / brand; personal memory covers workflows / projects / corrections / user.

This matters because it makes brand agents genuinely able to "carry" projects: tell Miora your brand positioning today—colors, fonts, auxiliary graphics, layout rules all land in project memory—and tomorrow a new session can continue VI extensions without repeating yourself.

A Real Workflow Sample

The original author used Miora to build a complete brand VI for his company. He fed in logo files, composition standards, color restrictions, font preferences (restrained, sharp, orderly, editorial), and concrete layout rules for slides, business cards, and event materials.

Miora (Max tier) produced a VI system in about ten minutes, plus PPT templates (dark and light versions), front/back business cards, and backdrop rules. A typical problem surfaced along the way: 45° angled cuts were imprecise—something AI can't yet solve, requiring manual vectorization in Illustrator. This exposes the "last mile" problem of design agents: AI can produce proposals, but standardized execution still needs humans.

The author's verdict: Miora has "WorkBuddy DNA"—models generate, Skills define process, memory provides continuity, and the canvas carries the output. Once Skill + project memory + brand rules + asset libraries are all connected, AI truly begins to understand "why a brand looks the way it does."

Comparison with WorkBuddy

• Shared traits: left-side column layout, skill marketplace, memory system, project-tracking design language.
• Where Miora is weaker: no native support yet for professional design formats (.psd, .ai, .svg)—a common gap among design agents that Miora plans to address.
• Where Miora is stronger: a denser out-of-box multimodal model matrix (image, video, UI, 3D all independently selectable), plus scene-based quick templates—brand design mode, film creative mode, e-commerce advertising mode, interactive game mode, web app mode—each with purpose-built templates.

Industry Implications

Design agents reaching the "harness stage" means the industry's competition has shifted from "model capability" back to "workflow packaging." WorkBuddy has already demonstrated: model weights are not the whole differentiator— whoever packages Skill + memory + multimodal combinations + scene templates into an out-of-box product wins.

The author's closing line: "In the future, a designer's most important skill may gradually shift from hands-on design—like a decade ago when I polished pixels in Photoshop—to what programming has become today: defining rules, building systems, judging proposals, controlling taste. Designing the harness for every project."

That captures the AI design battlefield for H2 2026. Coding agents have already transformed software development; the same revolution in design is beginning.

The open question: will Miora and WorkBuddy integrate? If so, what form will unified "design + coding" agents take? This is the product evolution most worth tracking over the next 3–6 months.

Source: Digital Life Kazk (数字生命卡兹克) WeChat article, 2026-07-22, by Kazk and Keda.