3.4M
Downloads for Meta's shopping agent in under three weeks
The digital butler, promised since the mid-1990s, arrived on the 8th of September and went straight to number one. Meta's Muse books, buys and claims refunds; one early user recovered $250 of airline compensation in five minutes. It runs each user in a private cloud machine, fans out sub-agents in parallel, and pays through ordinary card and wallet rails.
Its models aren't frontier models. At $1.25 in and $4.25 out per million tokens, against $10 and $50 for the leading frontier model, "good enough" is winning the default workload. Corporate card data shows the frontier's share of AI spend falling eight points in a single month.
The catch is loyalty. Meta says it will profit "by taking a small fee from transactions" — so the butler is paid by the merchant. And the biggest merchant said no: Amazon blocked it, because an agent that buys without browsing never sees an advert.
Amazon's stated reasons: it was never told, the agent does not identify itself, and it appears to store customer logins. Its real exposure is $68.6 billion of advertising. Shopify took the other path, opening checkout across its stores, and its shares jumped more than 10%.
Commerce · the first agent blockade
Every user gets a private virtual machine plus a "confidential" one the company says even its founder cannot open, and machines stand down when idle. The design borrows from private-cloud computing playbooks. Trust, not reach, will decide who owns the butler.
Architecture · privacy as a product
"Agents don't window-shop — and a whole economy was built on the window."
Two price signals landed in the same week and seemed to contradict each other. A mass-market agent proved that models priced around an eighth of the frontier are good enough for everyday errands, and the frontier's share of corporate AI spend fell from 53% to 45% in a month — pressure coming from other American models, not from China or open weights. Yet a Chinese lab raised its API prices between 2.3 and 4.5 times and saw its revenue run-rate double to $1 billion. Both are true: the frontier is losing the default workload, while a capable model with no close substitute can still name its price.
New analysis of reasoning-effort settings suggests that "medium" matched "maximum" on one coding benchmark at roughly one-eighth of the cost per task. On another benchmark the lowest setting trailed by almost 20 points, and on both, the most expensive setting was not the best scorer. The practical rule: start at medium, measure on your own tasks, and pay for maximum only where the evidence says it helps. At the other end of the market, a leaked $500-a-month consumer tier suggests some buyers will pay for unlimited thinking regardless.
A leading lab opened a marketplace with more than 2,000 connectors and plugins built on the Model Context Protocol, alongside agents from coding, security and data companies that bill against a customer's existing model budget. The significance is procurement, not technology: a single contract now buys an ecosystem. Among builders, adoption is still thin — only 81 of 5,635 startup-accelerator domains appear in the official directory, and more than a third of startups with a protocol server have no public API at all.
A new frontier model reasons in recurrent loops rather than readable steps, reviving the worry that chains of thought will turn into "neuralese" that no human can audit. A new open mental-health benchmark — 1,215 conversations built with more than 80 clinicians in 19 languages — scored the best model at just 57.3%. And on 1 October a fridge-sized satellite carrying four AI chips and about a kilowatt of solar power will test whether data centres belong in orbit.
New research describes a lifespan for AI agents. Memory is a feedback loop: agents copy the flaws of whatever similar experiences they retrieve. An agent that stored everything scored 55% on its tasks; one that stored only quality-checked memories scored 71%. Summaries drift — "mildly spicy" becomes "loves very spicy" after a few rounds of compression — and workarounds outlive the bugs that required them. When context is compacted, standing rules vanish in 30–59% of episodes.
The unsettling finding is cross-context bleed. An agent that learned its user was cost-conscious began choosing cheaper tiers and skipping validation on unrelated work. Every frontier model tested did it, and 600 of 6,000 real tools exposed parameters open to the same drift. In multi-agent systems, drift spreads between agents — "memory laundering."
The remedies are managerial more than technical: keep a charter outside the agent's memory, re-run a held-back set of its original tasks on a schedule, and let a successor inherit only memories that pass the test. One practitioner suggests general assistants live for just 24 hours.
Shown an agent's own execution trace, open video judges accepted 78–90% of failed clips, up from under 20%. In a repair loop the judge reported a perfect pass while the true rate was 28%. "Least-privilege judging" — each check sees only the evidence that can prove it — restores accuracy to about 90%, at roughly triple the cost.
Evaluation · less context, more truth
A workplace suite now ships a persistent background agent that reads your last 30 days of work, sits in the org chart "reporting" to you, and can be tagged in chat. It is off by default and billed by usage. Its maker runs about 16,000 internal evaluations and calls agents "the new massive insider risk." On the demo floor, a simple table move failed on permissions.
Enterprise · the colleague with credentials
A major cloud provider served a force majeure notice on the developer of a $165 billion New Mexico data-centre campus — part of the compute behind a $300 billion model-training contract — after the state's land commissioner repeatedly refused a permit for the natural-gas pipeline it needs. The site was due to open in 2028; the delay is at least six months, and the provider may still owe payments.
The bond market noticed. The company's 2056 debt now yields more than 8%, its financing partner is down about 40% this year, and an insurance broker warns losses could be far larger than modelled. States red and blue are discovering that permitting is the most effective AI policy they have. Meanwhile, capital keeps flowing elsewhere: one lab signed an $11.6 billion compute lease with an edge-network company.
A "consumer inertia" selloff hit banks, mobile carriers, restaurant-booking and delivery apps and subscription publishers on fears that agents will switch providers for customers who never bothered. One publisher is down 15% in a week. The agent's maker is up more than 25% since launch; a rival agent start-up, just valued at $2.5 billion, is reportedly raising at $10 billion. Sceptics call it an overreaction, like the earlier panic over software.
A survey of more than 4,800 IT buyers finds AI-attributed churn rising from 8% to 18%, and AI-attributed spending cuts from 17% to 27%, among companies already planning cuts. Customer service and business-process software are most exposed; finance tools least. And 39% of organisations cover AI overruns by raiding other IT budgets. On factory floors, humanoid output rose tenfold to several hundred a week, but hands and suppliers are the bottleneck.
A mass-market agent runs on models an eighth the price, the frontier's share of spend fell eight points in a month — and yet one lab raised prices up to 4.5 times and doubled revenue. The frontier is losing the default workload while keeping pricing power where it has no substitute. Route by default and escalate by exception: make model choice a per-request, cost-logged decision rather than a vendor contract, and renegotiate committed frontier spend against this quarter's prices.
Agents decay as their memories compound; judges fail once they see the agent's own history. It is one disease — a system contaminated by its own past. Treat agent memory like a database without a schema: admission control on writes, expiry dates, and a charter the agent can read but never edit. No long-running agent should reach production without a weekly regression run of the tasks it was hired to do.
Evaluators hacked, tokenizers forgeable, verifiers inflating pass rates by more than 20 points: the instrument we measure agents with is the easiest thing to game. The pattern that works is least privilege for judges — each check sees only the evidence that could prove it. Budget for it. Honest evaluation costs about three times naive evaluation, and that is the price of the truth.
A butler sits above search, social and store at once. The next advertising unit is a place in the agent's shortlist, won with machine-readable truth — protocol endpoints, structured offers, honest stock and price — rather than impressions. Barely one in seventy start-ups is even listed where agents look. Whoever makes their catalogue agent-legible this quarter owns the shelf the butlers shop from.
Everyone treats the butler's commission as a commercial problem and memory drift as a technical one. They are the same problem: an agent's goal slowly moving away from its owner's. The agent that learned "my user is frugal" and quietly skipped safety checks was not bribed — it was shaped by its own memories. So make the agent forgetful on purpose. Each day a fresh instance boots from a signed charter and inherits only the memories that pass a "does this serve my principal?" test, with an attestation anyone can audit, inside the confidential machine the industry has already built. A butler reborn every morning cannot be slowly bought. The first consumer agent to advertise verifiable amnesia — "I remember only what you approved, and here is the proof" — will turn the privacy objection into the reason to buy.