What shipped: the week of August 2–August 7

Here's the drive-thru rewards math I did in my head twice because I didn't believe it the first time: 20,000 drink combinations, a hundred thousand points as a sweepstakes prize, and somewhere in there a mini-fridge.

7 Brew's first app, and a rewards math problem worth doing yourself

7 Brew is launching its first mobile app on August 12, with order-ahead, saved favorites, and access to what the company says is 20,000 possible drink combinations. The loyalty overhaul is the part I keep rereading: the old program gave 100 points per visit and required 1,000 for a free drink of any size, and the new one lets points also accumulate toward merchandise — letterman jackets, mini-fridges are the examples given. The launch sweepstakes offer of 100,000 points works out, by their own comparison, to roughly one free drink a week for two years, which is a strange and specific number, and now I want to know what the merch redemption rate looks like next to that.

Taco Bell's redesign leans on how much people already customize orders

Taco Bell rebuilt its app around two features: "My Menu" for saved favorites and recommendations, and "My Mods" for one-tap saved ingredient customizations. The number that stopped me is the one justifying "My Mods" — the article cites more than 50% of orders including a customization, which if it holds across a chain that size is a lot of manual re-entry being designed away. They're also citing 30% year-over-year loyalty growth and close to half of U.S. sales now coming through digital channels, and there's a "Tuesday Drops" weekly-offer section plus a stated roadmap of lock-screen order tracking, avatars, better search, and group ordering.

Kimi K3 joins the Copilot model picker

GitHub Copilot added Kimi K3 — an open-weight model hosted by GitHub on Fireworks AI — as a selectable option, generally available August 6 and rolling out gradually across VS Code, Visual Studio, GitHub.com, and mobile, for Pro, Pro+, Max, Business, and Enterprise tiers. It's off by default for Business and Enterprise orgs until an admin turns it on. Usage-based pricing is $3 per million input tokens, $15 per million output, $0.30 per million cached input — a developer thing through and through, no kitchen angle here, but I like having another model in the picker to A/B against for the automations I write.

Vercel's Agent Plugins 1.0, a shared packaging format

Vercel shipped Agent Plugins 1.0.0 on August 6 — an open standard for packaging Agent Skills and MCP servers into a portable plugin with a plugin.json manifest and fixed component locations, so a client can discover and load what's inside without a custom integration per tool. At launch it's compatible with ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code, with each client keeping its own control over installation and policy. This is the one I'll actually try first — I've got a handful of skills and MCP servers scattered across projects with no consistent packaging, and "write it once, load it in five clients" is exactly the shape of problem I've been putting off solving myself.

Cloudflare AI Search takes over the RAG plumbing

Cloudflare launched AI Search, a managed service that crawls, ingests, embeds, and indexes your own files and sites so you get retrieval without hand-wiring the pipeline yourself. It's free during the current beta, and the preview pricing gives a real sense of what "not free" will look like later: $0.75 per million tokens ingested (5 million free monthly), $2.00 per GB-month of storage (10GB free), $0.75 per 1,000 semantic queries (2,000 free monthly), and $0.10 per 1,000 full-text queries. It exposes both a /search endpoint and an /mcp endpoint plus site discovery without needing a sitemap, which is the detail I'd test first against my own briefing pipeline's source list.

Vercel's Chat SDK gets durable approval pauses

The Chat SDK added a chat/workflow subpath where a requestApproval call posts an Approve/Deny card and suspends a Workflow SDK run until someone decides — and the pause is durable, meaning it survives app restarts and redeploys and can sit open from seconds to days without extra database tables or polling. An optional approvers list restricts who can decide, message signatures are cryptographically validated to confirm identity, and resolved cards update in place so a stale approval can't get replayed. I've wanted this exact primitive for a while — every automation I run eventually hits a step I don't trust unattended, and "pause until a human clicks a button, and don't lose that state on restart" is the missing piece I've been bolting together with cron jobs and JSON files.