Shift Lab

10/01/2026

The W3C's September working draft for WCAG 3.0 changes the shape of accessibility conformance, not just the checklist.

Where WCAG 2.x scores each criterion pass or fail toward a single A, AA, or AAA level, the new draft splits requirements into three tiers: core, supplemental, and assertions, the last of which lets an organization document its process rather than pass a binary test.

Notably, several provisions that were previously optional at the AAA level, including transcripts, audio descriptions, and plain-language summaries, are being folded into the mandatory core tier.

For any team maintaining a component library or design system, this is worth reading now rather than after it lands. We build accessibility into our design systems from day one, and we're already mapping our own components against where this standard is heading.

Read more: For Review: WCAG 3 Working Draft, September 2026
W3C Web Accessibility Initiative
(https://www.w3.org/WAI/news/2026-09-10/wcag3/)

09/30/2026

An article making the rounds this week lists what actually separates a white label development partnership that works from one that quietly falls apart: code ownership spelled out up front, a partner willing to work in your stack instead of insisting on theirs, real staging environments before anything reaches production, and every client conversation routed through the agency of record.

None of it is glamorous. All of it is the difference between a partnership a creative agency can build a practice on and one that becomes a liability the first time something breaks.

This is the model we've run for fourteen years: the technical partner behind the partner, invisible to the end client, accountable to the agency we're working with. The terms matter more than the pitch.

Read more: How White Label Web Development Lets Agencies Punch Above Their Weight
Nerdbot
(https://nerdbot.com/2026/09/19/how-white-label-web-development-lets-agencies-punch-above-their-weight/)

Plenty of marketing and design agencies land website projects they simply don't have the in-house development bandwidth

09/30/2026

Shift Lab is a 14 year old technical agency, and the way we work has changed even though what we are hasn't. The repetitive parts of product strategy, design, engineering, project management, and review now move through a structured, versioned pipeline of AI agent skills that every project pulls from one served system, rather than each project running its own static copy.

Human checkpoints still gate the work: spec sign-off, design review, architecture review, PR review, production sign-off, calibrated per project. The goal was never to remove judgment from delivery. It was to stop spending senior time on transcription so that judgment has somewhere to go.

We instrument the whole pipeline, which most agencies cannot say about their own process. Over the past 30 days it ran 861 times across five active client engagements, with 725 completing clean: a real number we can point to, not a claim we're making.

09/25/2026

A new model shipped on September 15 that only does one thing: it takes information in and returns a typed decision, a score, a yes or no, a category, with a confidence number attached. No conversation, no generated prose.

Within three days, Vercel, Cloudflare, LangChain and Langfuse had all integrated it. Vercel called it the fastest adopted model in its AI Gateway's history.

We build client architectures with the same instinct behind that adoption curve: match the tool to the task. A routing decision, a lead score, or a content classification step doesn't need a frontier language model reasoning through it end to end. It needs something fast, cheap, and accurate at that one job.

That's the kind of call we make on every stack we recommend, because Shift Lab has always been tech-agnostic. Not the biggest model available. The right one for what the system actually needs to decide.

Read more: Jev is the fastest-adopted model in AI Gateway history (Vercel) (https://vercel.com/blog/ai-gateway-jev-model-launch)

Within 24 hours of launching on AI Gateway, Jev from TypeSafe AI has been used by more than twice the share of teams of any other recent model launch in its first day.

09/24/2026

Design tokens have always had a translation problem. A color or spacing value gets defined in Figma, then re-typed or hand-converted to work in the codebase, the docs site, and whatever platform-specific format a mobile team needs. That gap is where handoff friction and "that's not what I meant" conversations live.

This month, the ecosystem closed a real piece of that gap. zeroheight adopted the DTCG token spec as its default export format, the same standard Figma, Sketch, Penpot and Tokens Studio already support, with Style Dictionary building it out into CSS, Tailwind, Swift and Compose.

We build component libraries and design systems on Tailwind and Storybook, and a shared token format means fewer manual conversions and fewer places for design intent to get lost between the file and the shipped product. Standardization like this is unglamorous. It's also exactly what makes a design system maintainable past launch.

Read more: What's new in the Design Tokens spec (zeroheight) (https://zeroheight.com/blog/whats-new-in-the-design-tokens-spec/)

The Design Tokens spec reaches its first stable version, introducing modern color spaces, groups/aliases, and token resolvers to make design systems scalable and cross-platform.

09/23/2026

GitHub announced this week that it's deprecating six Copilot models by October 19th, including several teams only started using a few months ago. New defaults are already swapped in behind the scenes.

This is becoming the normal cadence of AI tooling: models get retired faster than most engineering teams can build lasting habits around any single one of them.

It's part of why we've built our own AI-augmented delivery process around portability over lock-in. The structure that matters, the specs, the acceptance criteria, the review checkpoints, stays stable regardless of which model is doing the work underneath it. Models get swapped in and out. The process, and the human judgment reviewing what comes out of it, doesn't change.

That's the difference between a workflow that survives a vendor's roadmap and one that has to be rebuilt every time it changes.

Read more: Upcoming deprecation of selected GitHub Copilot models in mid-October (GitHub Changelog) (https://github.blog/changelog/2026-09-18-upcoming-deprecation-of-selected-github-copilot-models-in-mid-october/)

We will deprecate the following models across all GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions) on October 19th, 2026: Model Deprecation date…

09/23/2026

A legal analysis published this week made a point worth sitting with: most agency contracts were drafted before AI was part of daily delivery, and most haven't been updated since. Confidentiality clauses that never anticipated an AI tool as a "third party." IP assignments that assume a fully human author. Vendor agreements silent on what a subcontractor is actually running.

We've operated with documented internal standards for years, long before AI entered the picture, because predictable outcomes have always mattered more to us than moving fast and hoping. That discipline extends naturally to this: clear checkpoints on what's reviewed, by whom, and where responsibility sits.

Partners and clients don't need us to have all the answers on AI governance. They need a partner who already treats process, documentation and accountability as the default, not an afterthought bolted on once something goes wrong.

Read more: AI Is Reshaping Agency Work: Are Your Contracts Keeping Up? (O'Dwyer's PR News) (https://www.odwyerpr.com/story/public/25327/2026-09-18/ai-is-reshaping-agency-work-are-your-contracts-keeping-up.html)

Agencies are playing catch-up when it comes to drawing up the ground rules for AI. Getting up to speed starts with clearly defining how AI should be used.

09/21/2026

Shift Lab still runs the same core disciplines: product strategy, design, technology, project management, review. The shift is that the repetitive parts of each stage now move through a structured, versioned pipeline of AI agent skills, pulled live from one served system every project draws from, rather than each project running its own static copy.

That matters because routine work speeds up and gets more consistent, which frees senior people for architecture, judgment calls and the client relationship. Human checkpoints, spec sign-off, design review, PR review, production sign-off, still gate the work, calibrated to each project's maturity.

Most agencies can't tell you their own defect rate. We can, because we measure it. Bugs filed against the delivery model itself ran as high as 36 in a single week last month, and fell to 3 in the most recent full week, org-wide, with every hotspot fixed at the source.

09/18/2026

A useful stat from this year's enterprise AI surveys: 56% of companies now name a dedicated owner for what their AI agents are allowed to do, up from just 11% two years ago.

We'd argue that's the real dividing line between teams getting value from AI and teams stuck running pilots. An agent without a named owner and clear boundaries is a liability waiting to happen, not a productivity gain.

It's also how we've structured our own delivery model. Every stage, product strategy, design, technology, project management, build and review, has a deliberate human checkpoint attached to it. Oversight isn't bolted on after the fact. It's calibrated into the process from day one, lighter on mature projects, heavier on new ones.

Read more: The AI Agent Owner: The Role You Need in 2026 (Lusivision) (https://lusivision.com/en/blog/ai-agent-owner-role-2026)

More than half of enterprises now name an AI agent owner, and it is the single factor that separates agents that reach production from pilots that stall. Here is what the role is and how to fill it.

09/17/2026

Figma's latest push, an MCP server that lets AI coding agents read design tokens, components, and specs directly, is aimed at a problem we've lived in for years: the handoff between design intent and shipped code.

Coinbase's design system team wired this up through Code Connect and cut token costs by more than 20% while shaving 22% off task time, because the agent stopped guessing at what a component was supposed to be.

This is exactly the kind of tooling shift we track closely. Our product design and development teams work side by side on every engagement, component libraries, design tokens, and Figma files that map cleanly to production code. When the handoff is clean for humans, it turns out to be clean for agents too. That's not a coincidence.

Read more: LLM context design: Make your design system AI-ready (Figma) (https://www.figma.com/resource-library/llm-context-design/)