Across today's news, the question is who approves what an agent does: a board replacing a founder overnight, an agent that pays only after the owner signs, a sales team that kept a person on every message.
- Nvidia's sudden exit. Jensen Huang resigned as CEO effective immediately, and the CFO is now running Nvidia on an interim basis.
- A cheaper workhorse model. Google's Gemini 4 Hydrogen edges its larger Argon model on one index at $1.50/$7.50 per million tokens.
- Claude in the meeting. Claude can now join Google Meet calls, build documents live and act through attendees' tools with permission.
- Agents with a wallet. Replit's agent can now buy business supplies and launch Google Ads, with the owner approving every charge.
- Price the agent, not the feature. Jason Lemkin argues a real B2B agent should double ACV, while per-action fees on top of seats read as a price hike.
- The shrinking customer. Groove's founder closed signups at a profitable help desk after AI-driven seat cuts caused 59% of lost revenue.
- Signals beat volume. Needle found LinkedIn commenters on category posts replied 34.2% of the time, and a person still approves every message.
- Cap the AI bill. Two guides show how to cap unattended AI spend with card limits, prepaid credits and a quick tool audit.
This week, pick one agent workflow in your company and write down who signs off on its actions and its spending.
Developments
Business
Jensen Huang Resigns as Nvidia CEO, and the CFO Steps In While the Board Searches for a Successor (November 12, 2026)
Nvidia said on Thursday that Jensen Huang resigned as president, CEO and director, effective immediately. CFO Colette Kress is interim CEO while remaining CFO, and former Synopsys CEO Harvey Jones becomes the board's first chair. The release gives no reason, but it came days after Huang told staff he had been unfaithful, amid reports linking him to a woman US officials believe works for Chinese intelligence and to a relationship with a direct report. The board says it has no reason to believe company secrets were compromised.
So what
The lesson for most growth companies is governance, not chips: a 33-year founder-CEO was gone within four days, and the board had to stack two jobs on the CFO. Check that your board has a written emergency succession plan with a named interim and a conduct policy that explicitly covers the CEO. If you buy Nvidia capacity through a cloud, little changes this quarter, because the CUDA software ecosystem does not leave with him.
Confidence: Established fact — The transition is in Nvidia's release and SEC filing; the intelligence allegation is reported but not officially confirmed.
Go deeper Nvidia newsroom · The Neuron · The Deep View
Technology
Google's Gemini 4 Hydrogen Edges Its Larger Argon Model on One Index at $1.50/$7.50 per Million Tokens (November 12, 2026)
Google released Gemini 4 Hydrogen, a fast model for coding and agent work priced at $1.50 per million input tokens and $7.50 per million output tokens across its full 1-million-token context. It scores 51 on the Artificial Analysis Intelligence Index at its default setting, against 50 for Gemini 4 Argon, and 54 at a new xhigh thinking level. Google's Antigravity coding tool already switches between Carbon and Hydrogen in one session, which Google says came close to Carbon's results with 24% less quota and 19% faster completion.
So what
A frontier lab is now selling the plan-with-a-big-model, execute-with-a-cheap-one pattern as a product, which makes model routing a cost lever on the scale of a pricing decision. With Carbon's introductory price set to double after November 30, ask your engineering lead to test a two-tier setup on one agent workflow and compare cost per task at lower thinking levels before committing.
Confidence: Established fact — Launch, pricing and availability are on Google's blog; the benchmark and speed figures are Google's own.
Go deeper Google · TLDR AI · The Neuron
Product & Technology
Claude Joins Google Meet Calls as a Visible Participant, and Anthropic Offers a Self-Hosted Agent Platform (November 12, 2026)
Anthropic launched Claude Meetings in beta for Team and Enterprise plans. Claude joins Google Meet as a visible participant, takes notes, answers questions and builds docs and whiteboards during the call, and can use each attendee's connected tools, such as Slack, with that person's approval. It stays out of calls with outside guests by default and saves transcripts but not audio. Separately, a private beta of self-hosted Claude Managed Agents lets companies run the agent harness, sandboxes, credentials and session history in their own cloud.
So what
Meeting assistants are moving from note-taker to actor, which raises consent and permissions questions before it raises productivity. Before anyone turns on auto-join, set three rules: how attendees are told, whether external calls are allowed, and which connectors Claude may use. If you sell agents to regulated buyers, note the self-hosted option, because it answers the data-control objection that stalls many enterprise deals.
Confidence: Established fact — Both launches are documented in Anthropic's own product and help pages.
Go deeper Claude Help Center · Superhuman AI · TLDR AI
Customer Behavior
Replit's Agent Can Now Buy Business Supplies and Launch Google Ads, With the Owner Approving Every Charge (November 12, 2026)
Replit Agent can now browse a merchant's site, choose product options and prepare checkout for one-time business purchases, paying through Stripe's Link and Visa, for US Core and Pro users. The owner approves the final total, and during payment the agent loses shell access and general browsing, with approval checks enforced in code rather than left to the model. Pro users worldwide can also have it create Google Ads campaigns, including copy, images, budgets and conversion tracking, which launch only after the owner signs off.
So what
This extends yesterday's B2A thread: small businesses' agents are starting to shop for them. If you sell to SMBs, have someone test whether an agent can complete your checkout without a human rescue. If you are building agent actions into your own product, copy Replit's design choice of enforcing approvals in code, because a prompt instruction is not a control.
Confidence: Established fact — Capabilities and controls are from Replit's announcement; there is no usage data yet.
Go deeper Replit · TLDR Founders
Case Studies
AI Strategy & Transformation
Groove Closed Signups at a Profitable $4.7M-ARR Help Desk After Seat Downgrades Caused 59% of Lost Revenue (November 12, 2026)
Founder Alex Turnbull says Groove, profitable every year and never funded, had $4.7M ARR, more than 1,400 customers and 92% net revenue retention when he closed signups in March 2026. In 2025, seat downgrades caused 59% of lost revenue and happened three times as often as cancellations, as customers shrank their support teams. He moved the company to Helply, an AI support agent launched in January 2025 that now has 252 customers paying 1.3 times their previous amount, funded by Groove's cash flow.
So what
Customers did not leave Groove; they shrank. If you price per seat in a function AI is automating, run Turnbull's test: does your revenue fall when a customer's team gets smaller? Report downgrade-driven contraction separately from logo churn in your next board deck. His other lesson applies broadly: he says he hedged for more than a year when he should have committed.
Confidence: Informed perspective — The figures are the founder's own and unaudited, and he now sells the replacement product.
Go deeper Helply · TLDR Founders
Go-to-Market
Needle's LinkedIn Test: Commenters on Category Posts Replied 34.2% of the Time vs 9.9% for Page Followers, and a Person Approves Every Message (November 12, 2026)
Over two weeks in October, Needle ran 91 campaigns across six LinkedIn accounts to 2,253 new people, with AI finding buying signals and drafting each note and a person approving and sending it. People who had commented on posts about the category replied 34.2% of the time, event guest lists 32.2%, and followers of Needle's own page 9.9%. Messages triggered by funding rounds or SDR job posts drew no positive replies. The team tried to remove the human with prompts, eval loops and other models, and kept approval because the AI's judgment on which signal to use was weaker.
So what
Two lessons transfer. Interest shown in public conversation beat classic triggers like funding news, so ask your sales lead to report outbound reply rates by signal type. And like Clay earlier this week, Needle kept a human approval step after trying to automate it away. Note that these are replies, not booked meetings.
Confidence: Informed perspective — A vendor's own test with small samples per signal; Needle sells the tool it used.
Go deeper Needle · TLDR Founders
