A house newsletter · daily · researched and written by the Kurzia AI
Sunday 8 November 2026 · Issue 47 · 5 min read
METR inside Anthropic, AWS Kiro bots, Carta opens Evergreen, and 1,122 leaked Claude Code secrets
What's hot
The Verge · yesterday
METR's Anthropic report
Takeaway: around twelve times more productive
METR, an independent AI evaluation outfit, spent 11 weeks inside Anthropic looking at how much of the company's own research its models are now doing, and published on Thursday. Its verdict: Anthropic's models are improving about twice as fast as a year ago, and METR attributes most of that to AI assistance rather than to bigger training runs. Anthropic then blacked out much of the supporting evidence before the public version went out.
This is the headline your clients will have half-seen, and the useful framing is governance, not science fiction: the lab that lobbies hardest for caution is accelerating, and it alone decided what outsiders get to check. Cofounder Sam McCandlish said the detail was withheld because it is commercially sensitive and would hand rivals a roadmap. Polled internally, Anthropic's own staff put themselves at around twelve times more productive when working with the best in-house model, and three finished models are being held back — METR was told they are called Model 2, Model 3 and Model 4.
New tools & products
Kiro bots — named, long-running agents inside AWS's Kiro development environment, each keeping its own memory in a notes file you can open and edit. It is the vendor answer to the complaint you hear on every engagement — the agent forgets yesterday's correction — with tools and permissions scoped per bot and the human still accountable for the outcome. Bots are forbidden from handing work to other bots, deliberately, and each one starts with only the tools you give it at creation. The New Stack
“Developers end up as the go-between, carrying context from one session to the next.”
@GitHub in Slack and Teams — mention the bot in a Slack or Teams thread and it opens a coding session, files issues and raises pull requests; it is now generally available, with a Teams canvas for reviewing the plan and the diff in the conversation. This is where agents meet colleagues who never open an editor, so who may trigger a run stops being an engineering detail and becomes a policy you have to write down. It needs a paid Copilot plan and write access to every repo in the session, and anything it opens carries a line naming the chat and the person behind it. GitHub Changelog
Evergreen — a coding platform where colleagues and agents share the same cloud session, and the agents open their own pull requests rather than stopping when someone shuts a laptop. The pricing shape is the part worth stealing for your own cost models: no seat licences, model calls billed at provider rates, and a thin platform fee on top of every token. Carta's own operating numbers, not an outside audit: weekly merged pull requests up from roughly 1,300 in January to almost 20,000, with Evergreen charging $0.15 per million tokens and $0.35 per GB-month of storage. Carta
Claude Enterprise security center — an admin page that lines up a company's Claude settings against Anthropic's recommended configuration, alongside hooks that send prompts and tool calls to the data-loss tools a security team already runs. When a client asks how any of this gets governed, you can now point at shipped product rather than a custom build — and download the page as a report for auditors. The security center is in beta, and admins can run their own endpoint security sensor inside the virtual machine Claude and Cowork use on a user's computer. Anthropic
What people are saying
AgentConn · 2 days ago
The model/agent split
Takeaway: Anthropic leads at 57%
Traders on the prediction market Polymarket have pulled the two races apart: Google is the favourite for best model at the end of 2026, while Anthropic leads at 57% in the separate market for best agent. The argument doing the rounds is that a model produces raw capability, while an agent's reliability comes from the scaffolding around it — tool definitions, retries, memory, permissions.
It is a clean way to answer the client question you get every month, which is 'which model should we standardise on': the answer for a chatbot and the answer for autonomous code changes are allowed to be different vendors, and the second choice is stickier. Anthropic's Series G disclosure put Claude Code's run-rate revenue above $2.5 billion, more than half of it enterprise contracts.
“The AI industry is splitting into two supply chains. One produces intelligence (models). The other produces reliability (harnesses and agent infrastructure).”
News
Claude Code's "don't ask again" — picking the option that stops Claude Code asking about a command can save that command word for word into a project settings file — including any password or key typed into it. A config hygiene failure, not a model failure, and exactly the kind of thing to put on a client's pre-rollout checklist before anyone argues about which model to buy. Security firm Apiiro says 1,122 public copies of such files are sitting on GitHub. The Hacker News
ChatGPT connector routing — a developer reports that with two connectors — apps plugged into ChatGPT — pointing at the same backend, a follow-up tool call went out through the second one, carrying a one-time confirmation code the user had approved in the first. A crisp example for clients of why 'the agent asked permission' is not the same as 'the action happened where permission was given'. The server refused the call because the code was bound to the first connector's login; nothing told the user the switch had happened. OpenAI Developer Community
Grab — the ride-hailing company published how it keeps its mobile apps stable now that assistants have pushed up the volume of changes: static checks and tests before merge, an internal early-access build watched for crashes, then staged rollout and feature flags. This is the client-ready version of the speed argument — going faster is only safe if every stage can stop or switch off a bad change on its own. After release, crash monitoring in Firebase Crashlytics and BigQuery feeds automated triage that maps a crash back to the pull request that caused it. InfoQ
Gemini Omni 1.2 Flash — Google's video generation model is generally available, producing clips of up to 15 seconds, up from 10, and letting you adjust a clip across several chat turns instead of starting again. Marketing and brand teams will raise it, so know the price moved — and know that the Hacker News thread was mostly about Google's own mess of model naming, locked access and payment problems rather than the output quality. It costs $0.07 per second for 720p, a 30% cut from version 1.1, and every output carries an invisible SynthID watermark. Google
Muse Spaces — Meta's Muse assistant now builds app-like 'spaces' — a tracker, a planner, a dashboard — when you describe one, keeping them current from connected services, alongside a model update that is meant to hold on to what matters across conversations. Consumer territory, but this is precisely the 'describe it and it exists' demo an executive will show you before asking why internal delivery takes six weeks. On the October earnings call Mark Zuckerberg said the agents had 10 million users in the US and Canada less than two months after launch, TechCrunch reported. Meta AI Research
Kahn v. Anthropic — a US magistrate judge ruled that Anthropic must answer claims its Max plans were advertised as giving five or twenty times the usage of Pro when the weekly reality was lower, while dismissing several other claims. Subscription limits are now a contractual question in court, which is worth knowing the next time a client asks what their seat allowance actually guarantees. The plaintiffs calculate that for one model the $200 plan delivered about six times Pro's weekly usage, and the $100 plan 3.5 times. TNW
“Anthropic's per session disclosure is located two hyperlinks away from the primary sales pages on claude.com,”
Microsoft AI Frontiers — a founding leader writes that the Microsoft Research unit of roughly 50 people, whose work included the Autogen agent orchestration framework and the Fara computer-use agent, was absorbed into Microsoft AI in September, and that he and the other founding leads have now left. Autogen turns up in plenty of enterprise pilots, so it is worth knowing the team that created it has dispersed and that Microsoft's agent work now sits with a leadership the author describes as focused on scale over research. He says the move was made over the group's objections. hunch.net
“As far as I can tell we did good work and were non-consensually reorged away.”
One takeaway
Sell the plumbing review, not the model choice. Almost everything that shipped this week is control-plane work: memory with a permission list, admin consoles, per-token billing, attribution on a pull request. Almost everything that broke was plumbing too — a secret written into a config file, a call routed to the wrong app, a usage promise that did not survive a judge reading the sales page. Meanwhile the fastest-moving lab says it cannot show you its workings. Price the configuration review and the containment layer into your next engagement; nobody is going to pay for them separately. The Verge