Independent research essay, not peer-reviewed. The proposals are my own; every source is linked so you can check it.
Brand harnesses and brand hubs promise brands that AI agents can read and check. The history says living brand standards are old, and the research says agents receive only what can be written down, which is why the judgment in a brand still needs a person.
In the same week I saved two short videos selling the same future. In the first, the Brazilian designer Leo Becker describes a "brand harness": a brand built like software, "a living operating system" whose colors, typography, assets, strategy and governance are coded and served to AI agents and to people through protocols like MCP, a common standard for plugging AI tools into outside data (my translation from the Portuguese). In the second, Jack Watson presents "the future of brand guidelines", a "brand hub" that puts brand rules and coded tools together on one website, and warns that "if your guidelines are still sitting in a PDF, your work will genuinely get destroyed within a few months."
Both videos landed close to home, because I already run a small version of what they describe. The covers of this essay series come from a written guide with a fixed palette, fixed bans and a review checklist, and before Claude designs anything for me it reads my design rules and a taxonomy of visual techniques I built for it. So the question the videos raised for me was a practical one. When the reader of a brand guide is an AI agent, a program that acts on what it reads instead of just answering, what does it actually take from the guide, and what does it leave behind?
I checked the products, the people selling them and the research. The idea turned out to be older than the pitch, and the genuinely new part is the reader. An agent can read the part of a brand that can be written down and checked, and that part is getting cheaper to enforce. The parts that made brands hold, the judgment inside the rules, the way a company behaves and the shared language of the people who make the work, don't travel through a server, and the research gives reasons to think that a brand served only through its checkable layer will drift toward its checklist.
1. Two videos, checked
Both videos are sales, which doesn't make them wrong but does explain the urgency. Becker's Instagram bio describes a designer and "creative engineer" with an Adobe Creative Partner role, the video names Adobe first, and it ends by promoting Becker's newsletter, LBK Mag, and a paid recorded class on creative engineering (Becker 2026). Watson's company owns the product in the second video. Designer HQ belongs to Nonstop Branding Limited, the Manchester studio Watson founded, and its privacy policy says Designer HQ and the agency "are run by the same company" (Designer HQ). Designer HQ also pays creators half of the revenue from the studios they refer for a year, so other videos about it may be paid too.
MCP, the Model Context Protocol, is an open standard Anthropic released in November 2024 so that AI tools can reach outside data and actions through one shared format (Anthropic 2024). For a brand, an MCP server means that the same logo files, colors and rules can be handed to Claude, ChatGPT or the coding tool Cursor on request, instead of each tool working from its own copy of a PDF. Most of what the videos name runs on it, and most of it is real. Adobe announced a product called Brand Intelligence in April 2026, sold to large companies under contracts that include Adobe's own engineers to load each brand into it (Adobe 2026). Frontify, the Swiss brand-management platform founded in 2013, opened an MCP server with 52 tools in May 2026 (Frontify 2026). Encoded Brands, started in 2026, serves brands to agents over MCP and publishes its own open specification (Encoded Brands). The Brazilian service Becker calls "Rulis" is Ruler, a brand-governance studio launched in December 2025 (Propmark 2025). And Designer HQ does let Claude, ChatGPT or Cursor connect over MCP on every plan, including the free one (Designer HQ).
Some of it doesn't hold. I found only two uses of the exact phrase "brand harness", both from 2026: a startup called Defyner and Becker's own article, which says plainly that the idea isn't original. The word comes from AI engineering, where a "harness" is the scaffolding of tools, memory and checks around a model that lets it work as an agent. Mitchell Hashimoto started calling that work "harness engineering" in February 2026, and OpenAI used the term in a post a week later (Hashimoto 2026; Lopopolo 2026). Designer HQ is mostly a studio-management app for proposals, contracts and invoices, with brand hubs as one feature, and its only documented Figma connection is a plugin that exports selected frames. Nothing I found measures how fast PDF guidelines "destroy" anyone's work. And the most-quoted number in this market, that consistent branding raises revenue by 23%, is the average guess from a 2016 survey sponsored by the brand-template company Lucidpress, asked of respondents imagining their brand were always consistent. Nobody measured revenue (Demand Metric 2016). What's left after the checking is a real shift wrapped in two sales pitches, and the shift deserves a look on its own terms.
2. The brand book was never static
The pitch rests on a picture of the old brand book as a static PDF forgotten in a shared drive. The famous manuals of the 1970s were something else. The New York City Transit Authority's Graphics Standards Manual, made by Unimark with Massimo Vignelli and Bob Noorda in 1970, was a loose-leaf binder (Vignelli Center). NASA's manual, designed by Danne & Blackburn in 1975, was a ring binder whose index lists the pages "officially in effect as of January 1, 1976", which is version control done one page at a time (NASA 1976). When a rule changed, you replaced the page.
Those manuals were stricter than most brand books today. NASA's says the logotype "must not be re-drawn" and illustrates nine ways not to use it. It also asks for judgment that no checker could verify, since cover photographs should be "selected for their content, impact and graphic value." And strictness didn't produce consistency by itself. New York's manual specified Standard Medium rather than Helvetica, partly because of what the transit system's in-house sign shop could manage, and a coordinated signage system only arrived with capital spending from 1989 (Peterson 2011). NASA's leadership retired the "worm" logotype in 1992, whatever the manual said, and brought it back in 2020 (NASA 2020). Budgets and leaders decided what the binders could only describe.
Machine-readable brand values are older than agents, too. In February 2014 Salesforce published Theo, a "theme tokenizer" that turned one file of named values into code for websites, iOS and Android (Theo). Those named values became known as design tokens: a design decision stored as data under a name, such as color-text-primary set to black, so that every tool reads the same decision from one place. A community group at the W3C, the web's main standards body, published the first stable format for design tokens in October 2025, and the document is careful to say that it "is not a W3C Standard" (DTCG 2025). Digital brand portals such as Brandfolder, Frontify, Bynder and zeroheight date from 2012 to 2015. What changed in 2025 and 2026 is that these portals started answering software that acts on what it reads.
3. What the new reader can read
Becker is right that an agent struggles with a PDF, though the research describes a narrower problem than the video. Current models read the text in a PDF reasonably well and lose most of their accuracy on charts, diagrams and long files. On a 2024 benchmark of long, image-heavy documents, the best model scored 42.7% on a combined accuracy measure, and a 2026 study by Amazon researchers, not yet peer reviewed, brought its best setup to about 63% on the same benchmark, with extracted text alone doing worse than showing the model images of the relevant pages (Ma et al. 2024; Ak et al. 2026). A brand book keeps much of what matters in exactly those places: the clear-space diagram around a logo, the do and don't examples, the swatches. I found no study that compares a brand PDF with the same rules served over MCP, so the videos' comparison is plausible and untested.
What agents receive in practice is the layer of a brand that was already easy to write down. Anthropic's own example of a brand skill, the folder of instructions Claude loads to apply the company's look, contains seven colors, two fonts and rules such as "Cycles through orange, blue, and green accents", with nothing about voice, imagery or strategy (Anthropic 2025). Figma's MCP server, launched in beta in June 2025, does something smarter with the same layer. It hands the agent the name of a design variable instead of its raw value, because, as Figma puts it, "there might be many different tokens with the same red value" (Albaugh 2025).
That detail was the most useful thing I learned. A red with a name carries an intention: this is the warning color, or the accent for one product line. A red without a name is a hex code that any page can misuse. At its best, what an agent can take from a brand is a set of decisions that were already made and given names, which means the quality of the harness depends on the quality of the naming.
There is also a small sign of how fast the audience is changing. The Encoded Brands website carries text addressed to AI agents, asking them to read its brand file before anything else. Brands have started writing for a reader that isn't a person.
4. More rules, less brand
The harness pitch assumes that serving more of the brand to an agent makes its output more on-brand. I couldn't find a study that tests this directly, one that gives a model a brand guide and asks the brand's owners or its audience whether the output got closer. The nearest research points the other way.
When researchers gave models up to 500 instructions at once, the best models followed only 68% of them at the highest load and favored the instructions that came first (Jaroslawicz et al. 2025). A team at Adobe Research testing brand voice on small open models found that longer prompts cut instruction-following by 16 to 18 points, and that the models leaned toward "imitation rather than strict instruction adherence", copying the examples they were shown (Ramu et al. 2025). And a 2025 benchmark built around guidelines that may conflict with a model's common-sense defaults found that agents still have a long way to go in following them (Diao et al. 2025). Conflicting with defaults describes most brand rules, which exist to override them.
In my essay on the next default I argued that rules written to steer AI away from a look become a look of their own. These studies add a practical limit to serving "all of the brand strategy and governance" to an agent: past a point, more rules mean fewer of them followed, and the ones followed best are the earliest and the easiest to check.
5. The verification step
The second promise in Becker's video is verification: besides giving agents clean context, the harness checks what comes back. The tools that do this describe precise checks for what can be measured, such as brand colors, approved fonts, clear space around the logo and banned words. Several also claim to check tone of voice, but none of the vendors I read publish how often their verdicts agree with a brand team's. Their own documentation is modest about it. Marq, the brand-template company formerly called Lucidpress, says its Brand Guardian "provides guidance and flags issues, but it's advisory", and IntelligenceBank says "human reviewers remain in control of final sign-off at every stage" (Marq; IntelligenceBank 2026). Adobe's content-production software, GenStudio, scores brand compliance as the share of guidelines that passed (Adobe). Even Becker's article names measurement as the biggest open gap in the field.
Accessibility testing shows what happens when a checker becomes the judge. In 2019 the developer Manuel Matuzović "built the most inaccessible site possible with a perfect Lighthouse score", passing Google's automated website audit with a page that blocked keyboard use and hid its content from screen readers (Matuzović 2019). The UK Government Digital Service built a test page with 143 barriers and ran ten automated tools over it, and the best result found 41% of them (Duran 2017). A study of AI-built web pages found the opposite failure as well. After the pages were repaired to follow Material Design, Google's design system, about a fifth of the people shown both versions still preferred the originals, saying some issues, "like colorful themes, did not affect usability and were more engaging" (Yuan et al. 2025).
The anthropologist Marilyn Strathern gave an old economic warning its best-known wording in 1997: "When a measure becomes a target, it ceases to be a good measure" (Strathern 1997, p. 308). A brand score is exactly that kind of measure. Once a team works to raise it, the score stops telling anyone whether the work is on brand, and the brand starts to look like its checklist. That is the next default again, at the scale of a company.
6. Who controls a brand
If guidelines don't make a brand consistent on their own, something else does. The one academic study I found on brand consistency inside companies surveyed employees of 20 large Dutch organizations about the measures meant to keep their visual identity consistent, such as templates and access to up-to-date guidelines. The authors report that "Little correspondence was found between the use, perceived importance, and effectiveness" of those measures, and that managers setting an example mattered most (van den Bosch et al. 2004). Giving people access to the guidelines, in whatever format, wasn't what made them consistent.
Wally Olins, a founder of the branding firm Wolff Olins, put the control further out still: "brands are nothing like as powerful as they look. They are particularly vulnerable to us, the customers" (Olins 2003). What a company does, how it behaves and how people receive it decide the brand, and no server reaches any of those. Alla Kholmatova's book on design systems describes the inside version. At FutureLearn, designers documented patterns in a PDF brand document while developers built their own style guide, and neither became a shared language until the team changed how it worked together. The line I keep from that chapter is about names: "People enjoy using them and we don't need to enforce guidelines because they come with the name" (Kholmatova 2017).
So a harness governs what a company sends out, and that is useful. The things that make a brand hold sit somewhere else: in the example leaders set, in the language a team builds by working together, and in how the public receives the work.
7. My own harness
My own setup is a small test case. The covers for this series follow a written guide: a near-black ground, pale cream line work, one coral mark that appears exactly once and marks the essay's finding, no text, and a review checklist that includes looking at every render at thumbnail size. Before Claude designs, it reads my design rules, and for the covers it reads a taxonomy of visual techniques from my reference library, served to it over MCP, so that it chooses a technique I've named instead of a default.
The guide does its job on the letter. Renders come back in the right palette, with the coral used once and no stray text. The cover for my last essay shows where the letter stops. Its first render passed every written rule and was rejected because, at thumbnail size, its stacked diamonds read as the "layers" icon from design software. The rule that caught it was "no unwanted readings", which only a person can apply, and it worked because someone was looking.
There is a second lesson in the same setup, which I wrote about in The Next Default: kept long enough, my rules became a recognizable style of their own. What made the system useful was the named distinctions in it, which is Kholmatova's point about names and what Figma does with variables. In Unnamed, Not Obsolete I called the designer's part "taste with names attached", and a harness is one more place where that part has to be done by a person.
8. What I would build now
None of this argues against brand hubs. If I were handing a brand to a client today, I would ship its decisions in a form agents can read, because the letter of a brand should be cheap to apply: tokens with meaningful names, the logo files, the rules for type and color, and an MCP server if the client's team works with agents. Watson has a fair point here, too. Most businesses don't have a design team, and for them a hub that applies the basics correctly every time may be a real improvement over a PDF nobody opens.
I would also keep the judgment rules as judgment. "Selected for impact" was a person's call in 1975, and it still is, so the hub would say who makes that call and when. I would date every version the way NASA dated its pages. And I would judge whether the brand holds by how people respond to the work, which a pass rate can't tell me.
9. How this could be wrong
Three findings would change my mind. The first would be a study that hands a model a brand guide, has the brand's owners or audience judge the output, and finds it gets measurably closer to the brand. The second would be a brand checker that publishes how often its verdicts on tone and imagery agree with the people who own the brand, at a rate that holds up. That isn't far-fetched: on general chat quality, AI judges already agree with people about as often as people agree with each other (Zheng et al. 2023).
The third is the one I find most likely. For the many small businesses without a designer, the checkable layer may be most of what they need, and a hub that gets it right every time could matter more than anything I've said about judgment. If so, the harness is a good tool sold with the wrong history, and the judgment I'm defending belongs to brands that can afford a person to look. Either way, the brand now has a reader that does exactly what the guide says, so deciding what goes into the guide has become the design decision.
References
- Adobe. (2026, April 20). Adobe introduces Brand Intelligence and expands GenStudio content supply chain solution for customer experience orchestration [Press release]. https://news.adobe.com/news/2026/04/adobe-introduces-brand-intelligence
- Adobe. (2026). Adobe Brand Intelligence: Product description (effective October 2, 2026). https://helpx.adobe.com/legal/product-descriptions/adobe-brand-intelligence-base.html
- Adobe. (2026). Brand validation. GenStudio for Performance Marketing user guide. https://experienceleague.adobe.com/en/docs/genstudio-for-performance-marketing/user-guide/guidelines/brand-validation
- Ak, K. E., Mohta, J., Lee, G. G., Xu, Y., & Dimitriadis, D. (2026). An empirical study of VLM pipelines for long-document QA. arXiv:2609.29933. https://arxiv.org/abs/2609.29933
- Albaugh, J. (2025, June 4). Introducing our Dev Mode MCP server. Figma blog. https://www.figma.com/blog/introducing-figmas-dev-mode-mcp-server/
- Anthropic. (2024, November 25). Introducing the Model Context Protocol. https://www.anthropic.com/news/model-context-protocol
- Anthropic. (2025). brand-guidelines skill. anthropics/skills repository. https://github.com/anthropics/skills/tree/main/skills/brand-guidelines
- Becker, L. (2026, November 15). Brand Harness: Marcas como sistemas operacionais vivos e inteligentes. LBK Mag. https://blog.leobecker.com/brand-harness/
- Demand Metric. (2016). The impact of brand consistency [Benchmark report sponsored by Lucidpress]. https://d2slcw3kip6qmk.cloudfront.net/marketing/press/webinar/The-Impact-of-Brand-Consistency-Report-New_Cover.pdf
- Designer HQ (Nonstop Branding Limited). (2026). Homepage and privacy policy. https://designerhq.app/ ; https://designerhq.app/privacy
- Design Tokens Community Group. (2025, October 28). Design Tokens Format Module 2025.10 (Final Community Group Report). https://www.designtokens.org/tr/2025.10/format/
- Diao, L., et al. (2025). GuideBench: Benchmarking domain-oriented guideline following for LLM agents. ACL 2025. arXiv:2505.11368. https://arxiv.org/abs/2505.11368
- Duran, M. (2017, February 24). What we found when we tested tools on the world's least-accessible webpage. GOV.UK Accessibility blog. https://accessibility.blog.gov.uk/2017/02/24/what-we-found-when-we-tested-tools-on-the-worlds-least-accessible-webpage/
- Encoded Brands. (2026). Homepage and documentation. https://encodedbrands.ai/
- Frontify. (2026, May 5). Frontify MCP: Machine-readable brand intelligence. https://www.frontify.com/en/blog/frontify-mcp
- Hashimoto, M. (2026, February 5). My AI adoption journey. https://mitchellh.com/writing/my-ai-adoption-journey
- IntelligenceBank. (2026, May 19). Introducing AI-powered brand compliance for marketing content. https://intelligencebank.com/news/introducing-ai-powered-brand-compliance-for-marketing-content/
- Jaroslawicz, D., Whiting, B., Shah, P., & Maamari, K. (2025). How many instructions can LLMs follow at once? arXiv:2507.11538. https://arxiv.org/abs/2507.11538
- Kholmatova, A. (2017). Design systems, ch. 5, "Shared language". Smashing Media.
- Lopopolo, R. (2026, February 11). Harness engineering: Leveraging Codex in an agent-first world. OpenAI. https://openai.com/index/harness-engineering/
- Ma, Y., Zang, Y., Chen, L., et al. (2024). MMLongBench-Doc: Benchmarking long-context document understanding with visualizations. NeurIPS 2024 Datasets and Benchmarks Track. arXiv:2407.01523. https://arxiv.org/abs/2407.01523
- Marq. (n.d.). What is Brand Guardian in Marq. Help center. https://help.marq.com/what-is-brand-guardian-in-marq
- Matuzović, M. (2019, May 31). Building the most inaccessible site possible with a perfect Lighthouse score. https://www.matuzo.at/blog/building-the-most-inaccessible-site-possible-with-a-perfect-lighthouse-score/
- NASA. (1976; hosted edition revised c. 1989–1992). NASA graphics standards manual (NHB 1430.2). Danne & Blackburn. https://www.nasa.gov/wp-content/uploads/2015/01/nasa_graphics_manual_nhb_1430-2_jan_1976.pdf
- NASA. (2020, April 2). The worm is back. https://www.nasa.gov/general/the-worm-is-back
- Olins, W. (2003). On brand, ch. 11, "Brands: Who is really in charge". Thames & Hudson.
- Peterson, E. (2011, May 5). Book review: Helvetica and the New York City subway system. Urban Omnibus. https://urbanomnibus.net/2011/05/book-review-helvetica-and-the-new-york-city-subway-system/
- Propmark. (2025, December 1). Ruler Studio chega ao mercado com foco em governança de marca. https://propmark.com.br/agencias/ruler-studio-chega-ao-mercado-com-foco-em-governanca-de-marca/
- Ramu, P., Saxena, A., M Y, M., Sankar, V., & Basu, D. (2025). StyleAdaptedLM: Enhancing instruction following models with efficient stylistic transfer. arXiv:2507.18294. https://arxiv.org/abs/2507.18294
- Salesforce UX. (2014). Theo [Software repository]. https://github.com/salesforce-ux/theo
- Strathern, M. (1997). "Improving ratings": Audit in the British university system. European Review, 5(3), 305–321.
- van den Bosch, A. L. M., de Jong, M. D. T., & Elving, W. J. L. (2004). Managing corporate visual identity: Use and effects of organizational measures to support a consistent self-presentation. Public Relations Review, 30(2), 225–234. https://doi.org/10.1016/j.pubrev.2003.12.002
- Vignelli Center for Design Studies. (1970). New York City Transit Authority graphics standards manual, screenprinted binder [Google Arts & Culture record]. https://artsandculture.google.com/asset/new-york-city-transit-authority-graphics-standards-manual-vignelli-massimo/pgEC1D9xT-5X6A
- Watson, J. (2026, November 17). Is this the future of brand guidelines? [YouTube Short]. https://www.youtube.com/shorts/6iZ96dDXhW4
- Yuan, M., et al. (2025). DesignRepair: Dual-stream design guideline-aware frontend repair with large language models. ICSE 2025. arXiv:2411.01606. https://arxiv.org/abs/2411.01606
- Zheng, L., et al. (2023). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. NeurIPS 2023 Datasets and Benchmarks Track. arXiv:2306.05685. https://arxiv.org/abs/2306.05685
A note on sources: checked on November 19, 2026. Vendor pages (Adobe, Frontify, Encoded Brands, Designer HQ, Marq, IntelligenceBank) are evidence of what the vendors claim, not of what works. The NASA manual hosted by NASA as the 1976 edition is a later revision, with a letter signed by Administrator Richard Truly (1989–1992); the quoted rules appear in it. The New York manual's typeface comes from reviews of Paul Shaw's book, not from the manual itself. Ak et al. (2026) and Jaroslawicz et al. (2025) are preprints. The Becker quotes are my translations from the Portuguese video; the Watson quote is from the video's transcript.





