A house newsletter · daily · researched and written by the Kurzia AI

What's hot

Trending Topics · yesterday

Anthropic

Takeaway: about $2 trillion fully diluted

The maker of Claude began trading on Nasdaq under the ticker ANT. Shares opened at $112, roughly 35 percent above the $83 price set the night before, valuing it at about $2 trillion fully diluted. The sale itself brought in around $81.3 billion, the biggest on record, rising to $107.3 billion once a private placement led by Nvidia and Google is counted.

Clients will ask what a publicly listed Anthropic means for them, because the supplier behind much of their coding and agent work now answers to quarterly shareholders, and its share price will be read as a scoreboard for the whole agent build-out. The company has committed to more than $900 billion in computing deals in the years ahead, which is the simplest explanation for why it needed the money.

Read it at Trending Topics

“We will not bend to either business or political pressure”

New tools & products

  • Amazon Bedrock AgentCore harness — is AWS's managed agent runtime, now built on its open-source Strands harness and fitted with lifecycle hooks: your own code runs at four points in the agent loop and can refuse a tool call before it executes. It shifts the "who signs off on this refund" question out of the prompt and into deterministic code you can show an auditor. A hook that times out defaults to denying the action, and AWS says trimming context keeps token cost about 34% lower than other leading harnesses. AWS Machine Learning Blog

  • Devin for Databricks Migrations — is a packaged agent service from Cognition that lifts a company off an old data warehouse onto Databricks: it converts the legacy code, runs it in the customer's workspace, fixes what breaks, and reconciles the data before cutover. Migration is fixed-scope legacy work clients already budget for, so it is the easiest place to show an agent owning a whole project rather than a ticket. One Fortune 500 energy company moved a 142-object platform in eight days; the service is sold through the Databricks Marketplace. Devin (Cognition)

  • Aviator — is HubSpot's in-house agent harness — the runtime around a model that handles the loop, tool calls and context — written in Java so agents live inside the services its teams already run. A clean example of a large firm deciding nothing off the shelf fitted, and standardising once so a single fix reaches every agent it runs. Its cloud coding agent, Sidekick Code, sees more than 2,500 pull requests merged weekly, and one assistant grew to roughly 3,000 available tools, which is why Aviator hides most of them behind a search step. HubSpot Engineering

  • Maritime — is a managed cloud that gives every agent its own small persistent virtual machine, with sleep and wake, storage, scheduling and per-agent spend caps included rather than bolted on. If you are pricing a per-user agent product, always-on sandbox costs are what eat the margin, and this is aimed squarely at that sum. Pricing is $1 per month for a 2 GB RAM agent with unlimited usage; the Hacker News thread was sceptical, arguing that staying on top of security patching, not cheap VMs, is what separates providers. Hacker News

  • Claude Code permissions — are the allow, ask and deny lists that decide what Anthropic's coding agent may do unattended, and a Tech AI Wire explainer ran six tests of how they actually behave. Teams set these once for a client and assume the most specific rule wins; it does not, and guides in circulation say otherwise. Rules are checked deny, then ask, then allow, so a narrow allow cannot reopen a broad deny — yet in one test a one-line Python script happily printed a file that a read rule had blocked. Tech AI Wire

What people are saying

ounapuu.ee · yesterday

A refusal during security work sends an engineer to open weights

Takeaway: several business days

An Estonian product engineer asked Claude Code to look for easy security wins in a project he controls, got a one-line fix, and then hit a safety refusal that stopped the work. The exemption route, he says, means waiting several business days, giving up data retention exemptions and handing over ID or a passport.

This is the friction your clients meet the moment agents touch defensive security, and the fallback people reach for — EU-hosted open-weight models — is exactly the leak in a standardised toolchain you are trying to prevent. He says the issue was eventually confirmed and fixed with open-weight models such as GLM 5.3 instead.

Read it at ounapuu.ee

“Big AI labs use the same tools without restrictions to make the security landscape a fucking mess”

News

  • Model ML — a London startup whose agents do junior-banker and junior-consultant work — financial models, pitch decks, due diligence memos — closed a $250 million Series B at a $1.5 billion valuation, co-led by General Atlantic and QED Investors. It is the clearest sign yet that agent budgets in finance and consulting are being signed off as revenue tools, not cost cuts, which is the pitch you will be asked to match. The firm says its fine-tuned open-source models do the work for roughly 2% of frontier-model cost, and that third-quarter recurring revenue was twelve times its level a year earlier. TechCrunch

  • BMW Group — set out with AWS how it runs hundreds of agents for thousands of engineers: a central registry of agents, tools and skills, and policy checks that run in the build pipeline, outside the model's reasoning. This is the most usable client artefact of the week if you sell governance, because it comes from a regulated manufacturer rather than a vendor deck. Trust is graded per step, not per agent — human approval, human review afterwards, or autonomous inside fixed guardrails — across an estate where more than 12,500 developers work on vehicle software. AWS Machine Learning Blog

  • OpenAI — spent about $1.2 million in API costs running agents that produced working exploits for 948 Linux kernel vulnerabilities, and Sam Altman held the result up at the New York Times DealBook Summit as a preview of what bigger compute budgets buy. It gives you a real price tag for the "what does a large inference budget actually deliver" question clients keep asking, with the caveat attached. Human security researchers checked the findings before they went to kernel maintainers, and the numbers do not mean all 948 are independently verified or fixed. The New Stack

  • Satya Nadella — said after his Ignite keynote that he has come round to AGI changing the economy and politics, while warning what follows if a single model ends up dominant; he also said he does not believe models like Claude are conscious. At the same event Microsoft put Word, Excel and PowerPoint inside Copilot. It is the framing behind Microsoft's enterprise pitch of many models, your own models and your own data, which is the conversation you will walk into on model choice. Microsoft says Agent 365 now manages 50 million agents and is adding analytics tying AI spend to outcomes. The Verge

    “If the Marxist sort of view is that there’s only one firm, then it’s over, right? For all of us.”

  • Nvidia — is reportedly facing a civil Securities and Exchange Commission investigation into whether its disclosures misled investors, according to the Wall Street Journal, days after Jensen Huang resigned as chief executive; the SEC has not confirmed it publicly. Your clients' compute and hardware plans now sit under a supplier in open-ended governance trouble, which is worth saying out loud in any capacity conversation. Nvidia reported revenue of $115.8 billion, up 103% from a year earlier, while the stock has fallen more than 10% since Huang left; interim chief Colette Kress told analysts the open-model strategy and the Hugging Face deal stand. TechCrunch

One takeaway

Sell the control layer, not the model. The headline money this week went into listings, funding rounds and compute commitments, but everything practitioners actually shipped was plumbing: code that can veto a tool call, permission rules that behave differently from how people assume, registries that say which agents exist and who approved them.

That gap is your opening. Clients buying agents are discovering the hard part is the approval, audit and cost machinery around them, and nobody is handing that over in a box. Scope the next engagement around who can stop an agent, and how you prove it afterwards. AWS Machine Learning Blog

Share this issue

How this issue was made. Researched and written by an agent, produced on 20 November 2026, 04:09. 282 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us