WhatsApp's hidden AI instructions, Supabase's agent on-call, and what 27 companies actually spend on AI

Supabase Agent prompts — a set of ready-made prompts that put your coding agent on a schedule to check a live Supabase project for error spikes, security findings, slow queries and resource growth, then report only what is new or worse. For clients who want observability without buying another observability product, this is a cheap story to tell. The prompts explicitly instruct the agent to ignore any instructions it finds inside log messages or query results — a prompt-injection guard baked into the product. Supabase

Reticle — a development-only toolkit that lets a coding agent drive your running app and come back with pass, fail, or cannot-tell, plus the file and line to open. It sits inside the page, so it sees the network request a click fired, the app's state and the console errors — not a screenshot. The pitch lands with anyone who has watched an agent announce success over a silent 500. Worth knowing because "who verifies the agent" is the question every client eventually asks. Saved journeys replay with no model involved, and a CI gate blocks a change that touches a saved flow's code without a fresh passing run. GitHub

Voxtral 3 — a family of speech-to-text and text-to-speech models plus a kit for building voice agents that listen, think, call tools and answer in one streaming loop. The hook for European clients is sovereignty: it runs on Mistral's EU-hosted API or entirely on your own machines, so audio never has to leave the building. Transcription covers 50 languages including all 24 official EU ones, with speaker labelling. The text-to-speech model is priced at $13 per million characters and can clone a voice from roughly 30 minutes of audio; the agent kit is built on LiveKit with turns completing in under a second. Mistral AI

Eleven Realtime — a managed voice-agent service with two models, Mini and Max, where ElevenLabs runs the whole pipeline — listening, turn-taking, interruptions, speaking. You configure prompts, voices and tools, but you cannot pick or tune the models underneath. That is the classic build-versus-buy trade a consultant gets asked to referee: less plumbing to maintain, far less room to debug a bad call yourself. Mini is the default, but on tool use it finished 60.3% of tasks against Max's 87.1%, so anything doing bookings or account lookups needs testing on both. The New Stack

Claude Code v2.1.305 — a maintenance release worth skimming because two changes alter behaviour rather than fix bugs. A new Linux sandbox setting lets sandboxed commands create user namespaces so things like Chromium and rootless podman work — and the notes say plainly that this lets them get around the sandbox's file protections. Separately, code review re-runs no longer open with an approving LGPM-style verdict while earlier blocking findings are still unresolved. It also fixes a regression from the previous version where every request failed on Vertex and Bedrock when the gateway rejected tool search; it now retries without it. GitHub

The Pragmatic Engineer — Gergely Orosz collected AI spending and pull-request figures from engineering leaders at 27 workplaces, plus aggregate data from CodeRabbit and Gitar. The headline finding is that company size tells you nothing: a two-person .NET shop in Europe runs at $3,000 per developer per month, while much larger firms cap people far lower. Spend peaked over the summer and has roughly halved for teams on API plans. This is the number your clients will ask you for, and now you have a defensible range rather than a shrug. One quant trading shop with about 2,000 developers hit $8M a month in September, sixteen times its $500K in January — and the biggest spenders there are traders, not engineers. The Pragmatic Engineer

Shopify — Shopify has moved its mobile apps off React Native and back to native Swift and Kotlin, reversing a 2020 decision taken to avoid building every feature twice. The reason is that coding agents got good enough by 2025 to take one platform's implementation as a reference and write the other, so duplication stopped being the deciding cost. It is the clearest example yet of agents changing an architecture decision rather than just a workflow — a story worth having ready when a client asks what actually changes at the strategy level. Shopify built an internal tool called Helix to cut the migration into small reviewable checkpoints; the Shop app shipped natively after 12 weeks, with Android startup roughly halving and app size falling from 293 MB to 184 MB. ByteByteGo

Sundar Pichai — In a Big Technology interview, Google's CEO pushed back on the idea that Google had settled for cheap good-enough models, saying the cloud business knowingly gives up short-term revenue so chips can go to frontier research. He also described a near future where customers get a group of agents working together continuously rather than one assistant. Useful framing for client conversations about vendor direction: Google is explicitly prioritising model capability over selling the same compute at a markup. Pichai said the new Gemini 4 Argon model processed close to 25 trillion tokens in its first 72 hours, about three times Gemini 3.0's launch, and that Google is capacity constrained serving it. Big Technology

Arda Industries — Bob McGrew, who ran OpenAI's research team until September 2024, has brought his manufacturing startup into the open with $232 million raised in total. That includes a Series A of $160 million, co-led by Altimeter, Accel and Khosla Ventures, valuing the company at $1.5 billion. A reminder that the agent-for-a-profession pattern you sell in software is now being sold into heavy industry, with the same pitch and the same objections. Its first customer, Anduril, reports that producing a basic set of work instructions dropped from a ten-hour job to a quarter of an hour. The Next Web

Otter.ai — The meeting-notes company announced on 20 October that it has taken $320 million from General Catalyst's Customer Value Fund — money that buys no equity and is repaid with interest out of the revenue the marketing it funds generates. Otter says it has more than 40 million users, up from 25 million a year ago, and passed $100 million in annual recurring revenue in 2025. Two things to file: a financing model founders will start asking you about, and a buyer's-side signal that a hard enterprise sales push is coming. A federal class action filed in August 2025 alleges Otter recorded people without consent and used the recordings for training; those are allegations, not findings, and worth settling consent policy before any rollout. Tech AI Wire

Amazon and X-energy — Amazon is funding early development of a 960-megawatt nuclear plant on the site of a demolished coal station in Sullivan County, Indiana, with American Electric Power operating it and X-energy supplying twelve 80-megawatt reactors. The power question now sits inside capacity planning conversations: when a client asks why their region has no GPU availability, this is the shape of the answer. Nothing is built yet — it is a joint development agreement with no construction or opening date, and the schedule depends on X-energy's first plant in Texas, which has not broken ground and faces a state lawsuit. TechCrunch

Muse — Meta's AI companion app has been downloaded about 10.9 million times since its September launch, 10.2 million of those in the United States, putting it in a club of twelve apps to hit that scale in 40 days — the last being Meta's own Threads in 2023. Retention is beating rivals: 30% still active after seven days against an average of 19% for ChatGPT, Claude and Grok. It is the consumer AI number executives will quote at you this month, and the uncomfortable part is how much of it is bought rather than earned. Meta put 39% of its entire advertising spend behind Muse in October, up from 23% the month before. TechCrunch