A house newsletter · daily · researched and written by the Kurzia AI
Saturday 14 November 2026 · Issue 53 · 5 min read
ChatGPT Wallet lets agents check out; LangChain raises $300M; Anthropic prices its IPO this week
What's hot
TechCrunch · yesterday
ChatGPT Wallet
OpenAI has given its personal agents, called dots, a way to pay. You store a card once, the agent assembles the basket, and you sign off the checkout with Face ID or a fingerprint. The real card number is swapped for a one-use payment token, so neither the agent nor the shop ever sees it. It was built with American Express, Mastercard and Visa.
This is the question clients will bring you next: should our site accept purchases made by somebody else's agent, and who eats the fraud risk? The refusers are real — United said it will stop agents booking tickets and Amazon keeps them off its store — while OpenAI has published a playbook telling merchants how to wave the agents through. Actual demand is still tiny: during Halloween week, one of Uber's busiest of the year with about 320 million trips and orders, four rides came from an agent.
New tools & products
Brave Search API agentic payments — now lets an agent pay for each search as it makes it, in stablecoin, instead of you creating an account and managing an API key. It is the clearest working example yet of machines settling tiny bills on their own, and a different way to cap what a fleet of agents can spend. Each call costs $0.005 ($5 per 1,000); an unpaid request comes back with an HTTP 402 response quoting the price, and Brave is covering Base network gas fees for now. Brave
Google /swarm and Cursor multiplayer Projects — are rival answers to coding agents duplicating each other's work: Google's splits one developer's assignment across worker agents with a verifier checking them, while Cursor's lets different teammates' lead agents message each other. Clients ask about running fleets of agents; the usable answer today is that coordination holds up best inside one person's job. An Anthropic Fellows study found groups of per-user agents did worse than one agent serving everyone, by roughly 10 to 45 percentage points for Claude models. The New Stack
“The alpha from using swarms is ridiculous”
Self-hosted Claude Managed Agents — is a private beta that puts the coordinating server and the agent sandboxes inside a customer's own cloud or data centre, so credentials and session transcripts stay there. It answers the usual procurement objection, but the licence keeps the steering wheel at Anthropic: eight hours to disable or roll back a flagged feature, and no suppressing the telemetry sent back. The server arrives as a binary nobody may modify and may serve only Claude models; Spotify built its own in-house equivalent, Honk. The New Stack
“Self-hosted usually means the customer owns the change window. This license flips that.”
Building always-on agents — is Anthropic's engineering guide to agents that keep going for weeks, assembled from five parts: an inbox of incoming events, timers, memory, permissions, and an outbox for whatever reaches the user. It gives you plain vocabulary for the "can it just watch our systems and act?" conversation, plus a safe default to recommend: keep secrets out of the agent and let a proxy attach credentials to its outgoing calls. Anthropic
What people are saying
Gary Marcus · yesterday
Eli Lilly's chief pushes back on Amodei's ten-year cure claim
Takeaway: averaged 14 years
In a video tied to Anthropic's listing, the company's chief executive said cancer, Alzheimer's, heart disease and mental illness would be behind us in under a decade. Gary Marcus collected the response, which ran from oncology researchers to the chairman of Eli Lilly, and it was overwhelmingly sceptical.
You will be asked about this in meetings, and it sets the credibility baseline for everything else you tell a client about AI: when the ceiling is pitched this high, your sober claims get discounted with it. Having a pharma chief executive's own words to hand is more useful than arguing the science yourself. Marcus cites a count of the 20 most recently approved drugs, where the gap from first dose in a human to FDA approval averaged 14 years, with a median of 10.
“I think that would be amazing and I think it's not likely.”
News
Anthropic IPO — is at the pricing stage: the final share price is expected on 18 November and trading on Nasdaq under the ticker ANT is expected to start 19 November, with the US range estimated at $70 to $78. Clients read the valuation headline, so carry the underlying figures: $17.0 billion of Q3 revenue and a $23.3 billion net loss, most of it a non-cash charge. Simon Willison transcribed the roadshow video, in which Anthropic says it got Claude running on AMD chips in three days, almost unaided. Anthropic
LangChain — has raised $300 million at a $3 billion valuation to build out LangSmith, its platform for running, improving and governing agents. The cheque is for the control layer around agents rather than the models; in the same week Rox took $100 million at a $1.3 billion valuation for sales agents, and agent-security startup WhiteMatter $20 million. CapitalG led the round, taking LangChain's total raised to $460 million, with new LangSmith capabilities promised at a release event on 20 November. LangChain
Most large enterprises say agents are already live. A Genpact poll of 2,000-plus executives at billion-dollar companies found that 79% now run agents in production rather than pilots. Read it as the pressure your clients are under rather than proof of maturity: only a fifth use agents broadly across functions, so "in production" is carrying a lot of weight. Just 15% can trace the AI spend to a line in the profit and loss, and among the firms running agents, 92% report headcount flat or rising. Geta.Team
Agent Name Service — is a DNS-based way to verify that an agent is what it claims to be and to find it, launched formally by the Linux Foundation at KubeCon North America on 10 November with 11 organisations behind it. Agent identity is the first thing enterprise security teams ask about, and it now sits with a foundation rather than a single vendor. At the same event, Google handed Agent Substrate, which runs agent-written code in sandboxes on Kubernetes, to the CNCF. Geta.Team
Poppy 1.0 — is Sierra's standard for the moment an AI agent turns up at a company's website: the site states who the agent is acting for and which actions it may take, such as checking an order or booking a flight. Together with ChatGPT's wallet it is the other half of the same shift — your client's site needs a policy for admitting agents, not only for blocking them. Shopify has it on for every store; Mercury lets agents start payments that still clear the account's own approvals. The specification is still labelled a draft. TNW
Claude outage — ran twice on 13 November: first a sign-in failure on claude.ai and Claude Code, then hours of degraded service across the API, Claude Code, connectors, Managed Agents and Security. Both incidents are marked resolved. If a delivery plan assumes agents grind away unattended overnight, this is the failure mode that belongs in the risk column. Errors were back to normal levels at 15:33 UTC, but scheduled tasks and routines could run more than two hours late and some runs were skipped entirely. Claude Status
Agent-run attack tools multiplied after a takedown. Cloudflare says open-source kits that turn AI agents loose on a target and let them probe it have spread fast: in three days it saw more than 50 framework families scanning its customers. This is the agent-security question clients will actually raise this quarter, and Cloudflare's remedy is dull rather than clever: retire long-lived credentials and gate access by identity. Requests from these tools are up 30x since 1 October; in one incident a stolen, non-expiring CI/CD token drove 400 database dump requests in an hour. Cloudflare
One takeaway
Inventory who your client's agents are, and what they may spend. Several of this week's items are really one item: an agent now needs an identity, a spending limit and a permission list, and vendors are sprinting to sell all three. The model news, by contrast, was mostly financial.
So the next engagement opens with an inventory — which agents act for whom, which credentials never expire, who may approve a payment — before anyone argues about which model writes the code. And read the licence as carefully as the datasheet: rollback windows and telemetry clauses are now where the real constraints live. TechCrunch
