We answer on every line
Stop pinning the chatbot to the corner of the page
Vercel's Chat SDK picked up three new lines in one day: Notion comments, end-to-end encrypted XChat, and Slack Enterprise Grid. One codebase now answers wherever the client's people already type.
Today's art direction
The Answering Service
A Bass-era answering bureau brochure: one operator, every line.
Mid-century corporate design in the Bell System mold: a letterhead masthead under a double rule, flat offset-litho illustration plates, and typewriter labels set on engraved-looking panels. The palette is exchange-room hardware, pale seafoam paper and petrol ink with bell blue for links, and persimmon saved for live cords. The signature piece is a routing board that patches one jack through to every room the service answers in, with an honest capability note on each line.
Pages map to channels and comment threads to threads. No buttons or modals on this line; cards render as markdown.
End-to-end encrypted one-to-one and group chats. The adapter carries keys and signatures; streaming arrives as message edits.
One org-wide install across every workspace, with each tenant's caches kept apart.
Capabilities differ by line. The board says so on purpose.
The switchboard
Vercel spent Monday wiring new jacks into its Chat SDK: an agent built on it can now join Notion page comments, hold encrypted conversations on XChat, and install once across a whole Slack Enterprise Grid.
The pattern matters more than any single adapter. The same code already answers on Slack, Discord, GitHub, Teams, and WhatsApp; each new surface is configuration rather than a rebuild. For a working designer, the chat interface a client pays for stops being a widget pinned to one corner of the site and becomes a service that picks up wherever that client's customers already type.
The adapters are also refreshingly plain about their limits. Notion drops buttons and modals. XChat has no markdown, so tables get converted to plain text blocks and streaming happens through edits. Designing for the differences between lines is the new part of the job.
Tooling
Apple's new chips are pitched at the desk, not the cloud
Apple announced the M6 and M5 Ultra, its first 2-nanometer part and a quad-die flagship, landing in a new Mac mini and Mac Studio. The claims are aimed squarely at local models: a dual Neural Engine on the M6 and up to 4.5 times the peak GPU compute for AI of the M3 Ultra.
Real-user vitals for free
Vercel's Speed Insights added a free tier: 10,000 events per team every 30 days of performance data from real visitors rather than lab runs, on every plan.
Technique
JPEG XL is finally shipping
Mozilla published an intent to ship JPEG XL in Firefox, with Chrome moving in parallel and cross-browser support expected before the end of the year. For image-heavy pages the draws are progressive rendering, strong lossless work, and recompressing existing JPEGs with nothing thrown away. The decoder is a new Rust implementation Mozilla asked Google Research to build in place of the original C++.
Workflow
LinkedIn's agent keeps message slips
LinkedIn engineers walked through the memory system behind their hiring assistant: four layers covering the current conversation, recent activity, how a given user likes tasks done, and preferences aggregated across sessions. The stated goal is the answering bureau's oldest courtesy, never making a caller repeat themselves.
The write-up is a useful blueprint if the assistants you build keep asking clients the same intake questions every session.
Borrow this pattern
The routed coverage board
Replace the integrations logo wall with a routing diagram: one jack for the product, a trunk line, and a labeled cord out to every channel it serves. Each row gets the channel name, one plain sentence on what actually works there, and a filled status tag.
Use it on any services or integrations page where "works with everything" is the pitch. What keeps it honest is the capability note: naming what a channel cannot do (no buttons in Notion, no markdown on XChat) reads as confidence, and it saves the support thread later.
Prompt Lab
Build a one-page landing page for a telephone answering service, art-directed as a 1960s Bell-era corporate brochure. Canvas and palette: pale seafoam paper (#DEE7DE) with hairline vertical pinstripes about 84px apart; deep petrol ink (#1B3242) for text and panels; cream (#F4EBD3) for plates and cards; bell blue (#2A6395) for links; persimmon (#CF512B) reserved for cords, plugs, and NEW tags only. Type: a heavy grotesque display in all caps with tight tracking (Archivo 800/900) over a bookish serif body (PT Serif, 19 to 20px, line-height 1.65 or better), with typewriter labels in Courier Prime for every tag, caption, and form line. Masthead: a letterhead, not a boxed grid. Bold wordmark on the left, one typewriter meta line on the right, then a double rule (4px bar over a 1px hairline). Hero: two columns. Headline and serif deck on the left; on the right, an illustration plate matted on cream inside a 1.5px ink border with a typewriter caption underneath. Signature component: a routing board on a petrol panel. One labeled jack at top with a bullseye roundel, a vertical trunk line, and rows fanning off it through quarter-circle cord connectors that end in a plug dot. Each row: channel name in display caps, one plain sentence on what works on that line, and a filled status tag (NEW TODAY in deep persimmon, IN SERVICE as an outlined label). Use this instead of an integrations logo wall. Section markers: patch-bay labels. A small ink rectangle reading LINE 01, LINE 02 in typewriter caps beside each section title, with a hairline rule running to the margin. Close with a teletype message form as the contact block: a cream sheet with a ruled header, TO and RE fields in typewriter caps, and the message body in Courier. Guardrails: body text at least 19px, line-height at least 1.65, small labels at WCAG AA contrast on their actual background, at most one illustration style across all images (flat offset-litho spot inks with slight misregistration and paper grain), no readable text inside images, no gradients-on-dark, no glow.
Take it to Figma Make, to v0 or Lovable, or to Beaver Builder AI, which answers this line as written.
Smashing argued this morning that a website should never stop changing. The larger shift is that the site is no longer the only place your work picks up the phone.
Sources
- Bring your agent to Notion with Chat SDKVercel changelog
- Chat SDK now supports XChatVercel changelog
- Chat SDK now supports Slack Enterprise GridVercel changelog
- Apple introduces M6 and M5 UltraApple Newsroom
- Speed Insights adds a free tierVercel changelog
- Intent to Ship: JPEG XLMozilla Hacks
- Inside LinkedIn's cognitive memory agentStack Overflow blog
- Why Your Website Should Never Stop ChangingSmashing Magazine