This week in brief

Every week I go through what actually landed in my two inboxes — newsletters, daily briefs, funding notes, the occasional essay that turns out sharper than the commentary around it. Here’s what stood out between 10 and 17 November, and why I think it matters if you are running agentic AI in production.

A week with a hole punched through the middle of it. Jensen Huang resigned from Nvidia and the chief financial officer took the chair. Anthropic told prospective shareholders it has met the automated-research threshold in its own safety policy, days before its IPO. And Andrew Ng open-sourced an agent harness SDK and told his readers to go and build one, because interviewers are starting to ask. Three stories, one shape: everything you rent now has a single point of failure attached to it, and the part you build is the only part with your name on the door.

Nvidia lost its chief executive, and the stack felt it

I opened last week’s edition with a board investigation. This week it finished. On 12 November, TechCrunch reported that Jensen Huang resigned from Nvidia, days after admitting a romantic relationship with a woman US officials have alleged is a Chinese spy, and that Colette Kress, the company’s chief financial officer, has taken over as interim chief executive. The days around it were not quiet either. On 11 November, House Republicans gave Huang until the end of Wednesday to accept a closed-door briefing. By 13 November, Reuters had reported a third relationship, this one with a current Nvidia employee whose name it is withholding. On 16 November, President Trump said publicly that the affair did not threaten national security, contradicting his own intelligence officials, who TechCrunch reports concluded that Alice Huang is a Chinese spy.

Two details sit underneath the noise and matter more than it does. TechCrunch’s weekend edition noted that Nvidia’s security chief quietly left in August, and quoted a former intelligence officer describing this as exactly the kind of exposure that runs through everyday conversation and that most corporate security teams are badly equipped to catch. And on 13 November, separate research reported by TechCrunch found that China’s most advanced AI models were trained on export-banned Nvidia chips. Meanwhile a petition calling for Huang’s reinstatement, signed by Elon Musk among other technology executives, had passed 115,000 signatures by Saturday afternoon.

My take

I want to be careful here, because the gossip is doing most of the work in the coverage and almost none of the work in the analysis. Strip the scandal out and what you are left with is this: the single most load-bearing company in the AI supply chain changed chief executive in under a week, for a reason that had nothing to do with chips, demand, roadmap or execution. That is the definition of a risk you cannot engineer around. Nobody’s architecture review caught it. Nobody’s vendor questionnaire asked about it. And the export-control finding lands in the same week to make the matching point from the other direction — the controls everyone treated as a hard boundary turn out to have been porous the whole time. If your agentic programme has a line in its risk register that says “model availability” and nothing underneath it, this is the week to write the rest. Which provider, on whose silicon, under which jurisdiction, with what notice period, and what exactly happens to your production workload on the morning the answer changes. Ask it the ugly way: how many of your agents would stop working this week if one company’s supply or legal status changed, and could you tell me the number without opening a spreadsheet? I did this exercise on our own stack after reading Thursday’s brief, and it was more uncomfortable than I expected.

Anthropic says it crossed its own threshold, four days before the IPO

This is the story I have not been able to stop thinking about. In Chamath Palihapitiya’s 16 November newsletter, writing ahead of Anthropic’s IPO, he flags that on 10 November Anthropic told prospective shareholders that it has met the automated AI research and development threshold in its own safety policy. That threshold can be met two ways. Substitution: models can fully replace the company’s research scientists and engineers at competitive cost, within a factor of five. Acceleration: AI dramatically speeds up research, doubling the rate of progress in overall AI capability. The filing points at the second. Anthropic warned that its models may advance faster than its safety measures, and said it is still evaluating what that means.

It is not the only source saying so. Five days earlier, METR, a nonprofit that studies advanced AI systems, published an assessment after eleven weeks inside Anthropic. Most of its staff on the project believe Anthropic is improving its models at roughly twice the rate it managed between March 2024 and October 2025, with an uncertainty range of 1.3 to 3 times. The AI Futures Project, the forecasters behind the AI 2027 scenario, responded by saying that absent a deliberate slowdown, all of its authors now forecast a Superhuman AI Researcher by November 2027. The commercial numbers in the same newsletter are their own kind of vertigo: an annual revenue pace that went from about $1 billion at the end of 2024 to roughly $81 billion at the end of October; about $895 billion in commitments, almost all of it compute and data centres, including $44.6 billion with Nscale and $41.3 billion with Lambda; and $1.25 billion a month to SpaceX for its Colossus supercomputer, on a contract either side can cancel at 90 days’ notice after an initial period. On SemiAnalysis’s estimate, Anthropic kept about 65 cents of every dollar it charged for running models in the first half of 2026, having lost 94 cents on the dollar in 2024. TechCrunch reported on 10 November that all seven founders plan to sell the same number of shares, about $250 million each at the midpoint of the range, while keeping control.

My take

Read the safety disclosure as a buyer, not as a commentator. A vendor that a great many enterprises have standardised on has just told its incoming shareholders, in writing, that its capability is now outrunning its own controls, and that it has not finished working out the implications. I do not read that as alarmism — I read it as unusually honest, and I would rather have a supplier that says it than one that does not. But honesty does not change what it does to your plan. If the thing you are building on is improving at twice the rate it was eighteen months ago, then every assumption you baked in about model behaviour has a shorter shelf life than your procurement cycle. That is an argument for evaluations you own and run on every model change, not an argument for picking a different vendor. And notice the shape of the $895 billion: it is almost entirely rent. Compute leases, chip rentals, a supercomputer contract with a 90-day exit. The most valuable AI company of this cycle is building its business on commitments it does not own, which is precisely the position it is inviting its customers into. Rent the model. Own the harness. I did not expect the S-1 to be the best argument for it I would read all year.

Andrew Ng open-sourced an agent harness, and told everyone to go build one

The The Batch of 13 November opens with Andrew Ng writing that building an agent harness has become a rapidly growing method of building an AI agent, and announcing that the OpenWorker project has released the open-source OpenWorker Agent Harness SDK. His definition is the one I would use: a harness is the code that wraps around an LLM and gives it context, tools and guardrails, turning it into an application. His list of what a serious one needs is worth copying out, because it is a specification: context management, a memory architecture for long-running agents including compaction when the window fills, guardrails such as input and output checking, an approvals mechanism, security controls that keep credentials away from the model, and sandboxing.

Two lines in that letter made me put my coffee down. The first: “If you have not built an agent harness before, I encourage you to do so” — followed by the note that this skill is on the AI Engineering Skills Map and that job interviews increasingly ask people whether they have done it. The second is the reason he gives for open-sourcing it: unlike the harness tooling offered by model providers, this one is not beholden to any model, so you can build agents that run with any provider and on any cloud.

My take

I have been writing some version of “own the harness” in this series since August, and there is a specific feeling that comes with watching your argument stop being yours. This is the week the harness stopped being a position and became a job description. I am genuinely pleased about it, and I also think most people will take the wrong lesson. The lesson is not “go and adopt the OpenWorker SDK” — it is a good SDK and you should look at it, but an SDK is scaffolding, not an asset. The asset is what you put inside it: your context, your tool definitions, your approval rules, your evaluation suite, your action log. That content is the part a competitor cannot clone and a vendor cannot reprice. Ng’s own framing makes the point better than I have been making it: the reason to build the harness yourself is model independence. Everything a provider gives you for free is designed, perfectly rationally, to make the next model swap expensive. If you cannot change the model behind your agent with a configuration change, you do not have a harness. You have a dependency with good documentation.

The 948 exploits got a name, and the access list is the real document

Last week I wrote about an unreleased OpenAI model that found 948 flaws in the newest Linux kernel. This week The Batch named it and published the method. The model is GPT-6 Cyber, a version of GPT-6 Astra fine-tuned to find, exploit and patch security flaws, and in a mid-October scan it addressed 948 flaws allowing ordinary users to gain full control of a machine, at about $1,280 per vulnerability — against roughly $3,100 each for 166 such flaws when the same harness ran GPT-5.6 Sol in July.

What is new is the pipeline, and it is an agentic architecture diagram in prose. Scout agents identified security-relevant parts of the kernel reachable by ordinary users. Analyst agents read the candidates and flagged 2,308 flaws within 12 hours. A triage agent wrote a short program to trigger each flaw in a virtual machine, iterating until it fired or it ran out of approaches, confirming 1,724 flaws within 30 hours. A screening agent kept the ones with a plausible path to root. Exploit-writing agents produced 1,169 working exploits, and the pipeline checked nearby code for variants and wrote 948 distinct patches. Human security researchers reviewed and rated before anything went to Linux maintainers. The behavioural numbers are as striking as the capability ones: on a honeypot test measuring whether the model attacks an adjacent service it was not authorised to touch, the rate fell from 49.5% for GPT-5.6 Sol to 5.51% for GPT-5.6 Cyber to 0.07% for GPT-6 Cyber, and to nothing at all once OpenAI’s deployment safety instructions were in place.

My take

Skip past the exploit count and read the availability clause, because OpenAI has quietly published the best enterprise agent-governance checklist of the year. To get this model you must use single sign-on, multi-factor authentication, managed devices and least-privileged access, and restrict API requests to specific IP addresses. That is not a legal formality. That is a frontier lab stating, in operational terms, what it believes the minimum harness looks like around a genuinely capable agent. Hold your own agent deployment against those five controls and see how it scores. Most of the pilots I am shown would fail at least three. And the honeypot number is the one I would put on a slide for any board still debating autonomy: the difference between a 49.5% unauthorised-attack rate and zero was not a better model — it was training plus deployment instructions, which is to say the harness. Same capability, wildly different behaviour, entirely determined by the layer around it.

A 1990s keyword search matched the agent, and that should change a budget

The quietest item in The Batch was the one that should cause the most rethinking. Researchers at IIT Patna and TwoSpoon built READ — Reliable Embedding-free Agentic Document-search — which drops vector retrieval entirely and gives an agent three tools instead: a keyword search that returns matching line numbers, an outline tool that lists section headings with line numbers, and a reader that fetches specific line ranges. The agent loops: generate a search term, scan, check the hits against the outline, read from the table header down to the match, and refine when it gets too many hits, too few, or not enough context to interpret what it found.

They tested it on 51 questions about a 780-page government financial report, all systems running Gemini 2.5 Pro. READ answered 58.8% correctly. Chunk-and-embed retrieval managed 15.7%. Chunk-and-embed inside an agentic loop managed 27.5%. And BM25 — a keyword-ranking method from the 1990s with no model in the retrieval step at all — answered 51.0%, which the authors themselves deem statistically indistinguishable from READ. The Batch’s own comment is the sentence I would frame: before building an agent, it is worth finding out what a keyword search can do.

My take

I have sat in rooms where a six-figure vector database decision was made in under twenty minutes, on the assumption that embeddings are simply how retrieval works now. Here is a controlled comparison in which, on exactly the documents enterprises care most about — financial reports, contracts, regulatory filings, anything where a number means nothing without the unit label four lines above it — chunk-and-embed scored 15.7% and a method older than most of the people deploying it scored 51.0%. The failure is structural, not a tuning problem: chunking destroys exactly the relationship that gives the chunk its meaning. The cheapest thing on your roadmap this quarter is a baseline. Take fifty real questions your users actually ask, run them through plain keyword search over your real documents, and write the number on a wall before you procure anything. If the baseline is bad, you have justified the spend with evidence. If it is good, you have just saved a year. Either way you now own an evaluation set, which is worth more than the retrieval stack it was built to judge.

The ads are coming for your agents, not just your users

The Batch also ran the clearest account I have read of where the money in AI search actually is. Alphabet’s revenue rose 25% between July and September, but search revenue growth slowed to 12%, down from 19% and 17% in the previous two quarters — despite query volume growing at its fastest rate in five years. Google collected $63.35 billion from search and related ads in the quarter, and did not break out AI Mode, but an AI Mode query earns less than a comparable traditional search because it shows fewer ads and advertisers pay less for them. Ars Technica’s framing, quoted there, is that Google’s own models are cannibalising its search revenue. Google has now extended AI Mode ads to 11 countries, up from one. OpenAI shows ads in 89 countries to free and Go users and says its ad sales hit a $1 billion annual pace within 200 days — roughly $250 million a quarter against Google’s $63.35 billion. An OpenAI executive said ChatGPT’s weekly users grew from 1 billion to 1.2 billion in two months and that it handles about a fifth as many searches as Google, though Search Engine Journal notes OpenAI never said which Google figure it was comparing against, and Ahrefs put the number at 12% in February using a different definition.

The line that should concern anyone building agents is The Batch’s closing observation: the users seeing these ads will not necessarily be human. Agents shopping or researching on someone’s behalf will see them too, and Time magazine already sells ads aimed at agents rather than people.

Quick hits

  • Microsoft’s email outage became a full-blown global cloud crisis. A day-long incident that started with Outlook and Teams ended up taking much of Microsoft’s cloud offline (TechCrunch, 13 November). If your agents call Graph, that was your outage too.
  • Anthropic launched Claude Meetings, which drafts the plan as participants talk, maps a timeline on a whiteboard, answers questions in the meeting chat and pulls last year’s figures from connected work apps (TechCrunch, 11 November).
  • Databricks’ revised Genie agent tells chief executive Ali Ghodsi the unvarnished truth. At 12,000 employees he cannot know everyone, and reckons they are unlikely to confess what goes wrong; he uses the agent for that instead (TechCrunch, 11 November).
  • DeepMind’s AlphaGenome 2 reproduced in 30 minutes a gene-regulation finding that took a Francis Crick Institute team 18 months of experiments, per the researcher who led the original study, quoted in Chamath’s 16 November newsletter.
  • Dreamteam, founded by Freshworks alumni, raised $53 million for an AI-native CRM taking on Salesforce and HubSpot, with more than 100 customers across 13 countries (TechCrunch, 16 November).

The theme of the week

Dependency. Last week the through-line was permission — who is allowed to build, run and deploy. This week it narrowed to something sharper: every rented layer in this stack turns out to have a single point of failure, and none of them are technical. The chip supplier lost its chief executive over a personal relationship. The model vendor told its shareholders its capability is outrunning its own safety measures. A cloud provider’s email outage spread until much of its platform was down. Export controls that everyone treated as a hard boundary turn out to have been leaking restricted clusters for years.

Set against that, the capability layer had an excellent week: a named security model patching kernel flaws at $1,280 apiece, a genomics model collapsing eighteen months into half an hour, an open-source harness SDK anyone can pick up this afternoon. The capability keeps compounding and the dependencies keep concentrating. That is the asymmetry. The model, the compute, the chips and the cloud are rented from a small number of organisations whose stability is governed by boards, politics and personal conduct you will never have visibility into. The context, the harness, the evaluations, the controls and the record of truth cost a fraction as much, and are the only part of the stack that does not change hands while you sleep. Rent the model. Own the harness. This was the week the rented half spent proving the point on its own.

What this means if you’re deploying agentic AI

  • Count your single points of failure and put a number on each one. How many production agents stop if one model provider has an outage, a legal problem or a leadership crisis? How many stop if one cloud does? You should be able to answer both from a record, today, not from memory.
  • Score your deployment against OpenAI’s own access list. Single sign-on, multi-factor authentication, managed devices, least-privileged access, IP-restricted API calls. A frontier lab decided those five are the minimum around a capable agent. Run the comparison honestly and fix whatever fails.
  • Baseline your retrieval with keyword search before you buy anything else. Fifty real questions, plain search, write down the score. On exactly the document types enterprises care about, a 1990s method matched an agentic one and beat chunk-and-embed by more than three to one.
  • Treat evaluations as the thing that absorbs model change. If a vendor is now improving models at twice its previous rate, the only sustainable response is a suite you own, run on every version bump, that tells you in an afternoon whether behaviour you depend on has moved.
  • Make the model boundary real and then prove it. Not “we could swap providers in principle”. Actually swap one, in a non-production environment, this quarter, and time it. The number you get is your true switching cost, and it is probably not the one in your deck.
  • Decide now what your agents are allowed to be sold. Ads aimed at agents already exist. If an agent of yours researches suppliers, compares products or books anything, you need a position on what it is permitted to act on and what gets escalated to a person.

Frequently asked questions

Jensen Huang resigned from Nvidia — what should enterprises running agentic AI actually do about it?

On 12 November, TechCrunch reported that Jensen Huang resigned as chief executive of Nvidia and that chief financial officer Colette Kress took over on an interim basis, days after he admitted a relationship with a woman US officials allege is a Chinese spy. Reuters subsequently reported a third relationship involving a current Nvidia employee, House Republicans had called him in for a closed-door briefing on 11 November, and on 16 November President Trump said the affair posed no national security threat, contradicting his own intelligence officials. A petition for his reinstatement signed by Elon Musk and other technology executives passed 115,000 signatures. TechCrunch also reported that Nvidia’s security chief left quietly in August, and that separate research found China’s most advanced AI models were trained on export-banned Nvidia chips. The practical enterprise response is not to change supplier; it is to make supply dependency an explicit, quantified line in the risk register. Record which model providers your production agents call, which silicon and jurisdiction sit beneath them, what notice period you would get if availability or legal status changed, and how many agents would stop. Then make the model boundary in your harness real enough that swapping a provider is a configuration change rather than a re-platforming project, and test that claim in a non-production environment rather than asserting it.

Anthropic says it met its automated AI R&D threshold — does that change how enterprises should plan?

Ahead of its IPO, Anthropic told prospective shareholders on 10 November that it has met the automated research and development threshold in its own safety policy, which can be satisfied either by substitution (models fully replacing its research scientists and engineers at within five times the cost) or by acceleration (dramatically speeding up research and doubling the rate of progress in overall AI capability). The filing points at acceleration, and Anthropic warned that its models may advance faster than its safety measures and that it is still evaluating the implications. METR, after eleven weeks inside the company, assessed that Anthropic is improving models at roughly twice the rate it managed between March 2024 and October 2025, with a range of 1.3 to 3 times, and the AI Futures Project responded that all its authors now forecast a Superhuman AI Researcher by November 2027. For enterprises the planning implication is about shelf life rather than vendor choice: if model behaviour is changing at twice the previous rate, assumptions encoded in prompts, tool definitions and downstream logic will expire faster than a typical procurement cycle. The durable response is an evaluation suite you own and run on every model change, plus a harness with a clean model boundary. All figures here come via Chamath Palihapitiya’s 16 November newsletter, citing Anthropic’s own disclosures, METR, the AI Futures Project and SemiAnalysis.

A 1990s keyword method matched an agentic retrieval system — should enterprises rethink RAG?

Researchers at IIT Patna and TwoSpoon built READ, an embedding-free agentic document-search method that gives a model keyword search, document outline and line-range reading tools instead of vector retrieval, and tested it on 51 questions about a 780-page government financial report with all systems running Gemini 2.5 Pro. READ answered 58.8% correctly, chunk-and-embed retrieval 15.7%, chunk-and-embed inside an agentic loop 27.5%, and BM25 — a keyword-ranking method from the 1990s with no model in the retrieval step — 51.0%, which the authors regard as statistically indistinguishable from READ. The failure of chunk-and-embed here is structural rather than a tuning problem: in financial reports, contracts and regulatory filings, the meaning of a passage frequently depends on a unit label, date or definition that sits outside the chunk. The recommendation is not to abandon vector retrieval universally, but to stop treating it as the default. Before committing to a retrieval stack, run fifty real user questions through plain keyword search over your real documents and record the score as a baseline. The exercise costs very little, produces a reusable evaluation set, and will either justify the investment with evidence or save a year of work. Reported in The Batch, 13 November 2026.


From Veehive Labs

Fourteen weeks in, and this was the week the rented half of the stack made the argument for me. Rent the model, the compute and the chips; own the context, the harness, the evaluation, the controls and the record of truth. In seven days, the chip supplier lost its chief executive for reasons no architecture review would ever have surfaced; the model vendor most enterprises have standardised on told its incoming shareholders that its capability is outrunning its own safety measures; a cloud provider’s email fault spread until much of its platform was offline; and export controls that everyone treated as a hard boundary turned out to have been leaking restricted clusters for years. Over the same seven days the capability itself got cheaper and better again, and Andrew Ng open-sourced an agent harness SDK and told his readers that interviewers are now asking whether they have built one. Concentrated dependencies, compounding capability. The only layer that is neither rented nor contested is the one you build. Veehive Labs is a Dubai-based AI innovation lab and custom AI product development company, building model-independent enterprise AI harnesses, custom agents, RAG pipelines and sovereign, private-AI deployments for regulated and operationally complex organisations across the UAE, KSA, GCC and beyond.

If you’re moving from AI strategy to production, start with a focused AI Discovery Sprint — we map your use case, data, systems, agent inventory, permissions model, context retention, evaluation suite, model-swap path, supply dependencies, cost ceilings, jurisdiction assumptions, risks and delivery architecture before the build begins.

Build Your Organisation’s AI Capability

Start with a focused AI Discovery Sprint. We will map your use case, organisational knowledge, systems, MCP integrations, security requirements and delivery architecture.

Agentic AI This Week is written by Sathish Jeyakumar, Founder & CEO of Veehive. Bring your preferred model — we make it understand, speak and operate like your organisation.

← All insights