Jensen Huang is out at Nvidia, days before earnings. And this week, agents got cheaper to run and easier to sell to.

Today's haul: 16 tools · 4 resources · 20 reads · 11 numbers · 16 things that happened. Every tool, resource and read below has a working link. No link, no listing.

The 60-second catch-up

Nvidia lost its founder-CEO overnight, and the board named a stand-in the same day.

Jensen Huang resigned as Nvidia's president and CEO and left its board, effective immediately. CFO Colette Kress is interim CEO and stays CFO. Director Harvey Jones becomes the company's first chair (Nvidia, SEC 8-K). Earnings are still set for Tuesday 17 November. Inside the company, an internal post asking the board to bring him back has drawn more than 1,600 comments (Wired). The lesson for a small company is not about scandal. It is that the person who can sign for the company when the founder can't was named in one filing, on one day.

→ do this: write down who signs payroll, contracts and the bank mandate if you are out tomorrow, and tell that person.

Agents got cheaper this week, from better models and from reading less.

OpenRouter ran one task through GitHub's MCP server and through the gh CLI on the same model. The MCP route used 59,417 input tokens against 1,421 (OpenRouter, which sells tooling for this). At incident.io, agent spend fell from about $4,500 to about $1,600 per engineer a month after a model upgrade, while releases went from 35 a day to 140 (Gergely Orosz, who invests in the company). Google priced Gemini 4 Hydrogen, its fast coding model, at $1.50 in and $7.50 out per million tokens across the full 1M context (Google).

→ do this: log input tokens per task for one agent workflow this week. If a tool call is eating them, try the CLI version before you change models.

Agents are starting to buy, search and screen deals on their own. Write for them.

Brave's Search API now takes per-search payment from agents in USDC, at $5 per 1,000 searches, with no account or API key (Brave). a16z's Olivia Moore calls this the B2A economy, where agents are the customer (a16z). Haseeb Qureshi says founders have started sending investors a company.md, a fundraising brief written for the investor's AI agent to read (Haseeb Qureshi). Two OpenAI engineers proposed a draft standard for agents to pass a user's country, time zone and locale to websites. It has not been adopted (Joost de Valk).

→ do this: write a one-page company.md with what you do, your numbers and your ask, and put it next to your deck.

Tools

Build & ship

Rust 1.100 — The never type ! is finally stable. Cargo gets dependency cooldowns (min-publish-age), so a freshly published crate version can't land in your build until it has aged. · Open source (MIT/Apache-2.0)
Website to Markdown API — Send a URL, get the page back as Markdown, including pages that need JavaScript to render. · Free plan with 200 credits/mo · Scale $149/mo
Facade Code — Runs Claude Code, Codex and other coding agents side by side inside VS Code. · Open source (MIT)

AI & agents

⭐ Gemini 4 Hydrogen — Google's fast model for everyday coding and agent work. In Antigravity, the new Orchestrator lets Gemini 4 Carbon plan while Hydrogen makes the edits in one session. · $1.50 / $7.50 per 1M tokens · Orchestrator on paid AI Pro and Ultra
Nod — n8n's open-source personal agent. Bring a ChatGPT Plus or Pro login or any AI key. Research preview. · Free · open source (Apache-2.0) · provider fees apply
AgentRun — A self-hosted layer around your agent. It keeps what each run learned, survives restarts and caps spend. Beta. · Open source (Apache-2.0)
Kapso — Give it a goal and it runs experiments toward it, learning from each run. Open-sourced this week. · Open source (MIT)
A2UI 1.0 — Google's open protocol for agents that reply with interactive screens drawn in your app's own components. Web is 1.0; the iOS and Android renderers are still 0.9. · Open source (Apache-2.0)
Scribe v3 — ElevenLabs' new speech-to-text model, with a live version. ElevenLabs says it makes 27% fewer word errors than v2 across 102 languages. · No price stated
RealtimeVAD — Tells a voice app the speaker has stopped, a median 100 ms sooner than Silero, the usual speech detector. Alpha. · Open source (MIT)
Edge0-145B — A 145B-parameter model built to run on a laptop. The memory and speed figures are the company's own. · Open weights · Edge0 Community Model License
smolmaxxing — A patched llama.cpp that runs 16B models on a TV box and a 2013 iPad. · Open source (MIT)
Runway Praxis-1 — A 27B open model that controls robots by predicting video and actions together. · Open weights (Apache-2.0)

Design & create

zigzag — A free embroidery editor that turns SVG art into stitches and exports machine formats. Built because a $300 machine wanted a subscription. It has an MCP server, so an agent can design too. · Open source (Apache-2.0)

Growth & ops

⭐ Brave Search API, agent payments — Agents pay per search in USDC on Base through the x402 protocol. No account, no API key. · $5 per 1,000 searches
Databricks Genie Agents — Agents that answer questions across your tables, documents and tools such as Jira and Slack, and can act on the answer. Mostly preview or beta. · No price stated

Resources

Benchmarks you can re-run

SaaS renewal price increases by vendor, 2026: Cledara's table of what vendors raised at renewal and the contract clauses that let them. Start talks 90 days out.
MCP vs CLI vs Skills: what each costs in context: OpenRouter measures the tokens each approach burns on the same task. Run the same comparison on the tools your agents call most.

Learn the craft

Where the disabled button got its rule: How eight design systems handle disabled buttons, traced back to the Apple Lisa.
What is eval-driven development?: LangChain's how-to on using evals to decide every change to an agent. A vendor guide, but a clear one.

Reads

☕ Under 5 minutes

Company.md: a fundraising brief for investor agents: Haseeb Qureshi on founders writing a pitch for the investor's AI to read first.
The fall of what you can do and the rise of who you know: Jason Wei on why connections matter more once AI makes skills cheap.
Our software engineering interviews are changing a little: Jane Street is trialling rounds on design, reviewing code and working with an LLM.
Two kinds of companies to build right now: Aditya Agarwal says build either companies that spread AI into the economy or companies that speed up science.
3-2-1 for 12 November: James Clear's short notes, including an Artemis II astronaut's question about making the first right choice.

🍵 5–10 minutes

Details of incident.io's AI spend: Gergely Orosz, with data from incident.io's Lawrence Jones: spend per engineer, the share of PRs with agent code, and why output per person rose less than releases did. Orosz invests in the company and says so.
Welcome to the B2A economy: Olivia Moore on businesses that sell to agents rather than to people.
Who captures the most margin in AI?: Tomasz Tunguz models the AI value chain in dollars per watt.
Flo's long-tail subscriptions: Flo's Mike Charniatsou on why 40% of its subscriptions start more than six months after install, and how it keeps it that way.
How to stay current, get hired, and last in AI: Nikhyl Singhal asks three AI product leaders how they keep up without burning out.
AI roleplay apps: Dru Riley's report on companion apps: the Korean playbooks, 18+ ratings and App Store exits.
AI is turning Apple's app review into a nightmare for synth developers: iOS music apps stuck in review for weeks. Developers suspect AI reviewers; Apple has not said.
Agents sharing user signals: Joost de Valk on the OpenAI engineers' draft for agents to tell sites where you are.
How two side projects became Dropbox's agent: Dropbox on how two engineers' experiments became the agent that acts inside the product.

📚 Longer, worth it

Refounding Ro to a $3B+ run rate: Ro's CEO Zach Reitano on rebuilding the company with fewer people than it had in 2021.
Why should managers exist?: Sachin Benny on why AI has not yet made companies smaller.
Escaping SPACE, part II: Perplexity tried to get models out of its sandbox. None escaped, and the specialist cyber models did worse than general ones.
Shopify's generative discovery model: Shopify on running search and recommendations from one generative model.
Checkpoint forensics: Will DePue finds traces of Nvidia Blackwell chips in six recent Chinese open models, and says that shows access, not smuggling.
Turning a 22-year-old Postgres bug into code execution: Mehmet Ince's write-up of the bug behind this week's Postgres patch.

Numbers

$17.0B: Anthropic's third-quarter revenue, more than 12 times a year earlier (Anthropic IPO site).
59,417 vs 1,421: input tokens for one task through GitHub's MCP server against the gh CLI (OpenRouter).
$4,500 → $1,600: incident.io's monthly AI spend per engineer, before and after a model upgrade (Gergely Orosz, Lawrence Jones).
35 → 140: incident.io's releases per day, as engineering grew from 18 to 50 people (Gergely Orosz, Lawrence Jones).
8–12%: average yearly SaaS price rise at renewal, and 15–25% from aggressive vendors (Cledara).
49% vs 19%: savings when renewal talks start 90+ days out, against 30–90 days (Cledara).
27%: AI-native tools' share of software spend across Cledara's customers, up from about 3% (Cledara).
40%: Flo subscriptions that start more than six months after install (Mike Charniatsou, Flo).
463 → 406: Ro's headcount while its run rate went from $270M to more than $3B, by the CEO's own count (Zach Reitano).
52: security vulnerabilities fixed in Thursday's PostgreSQL releases (PostgreSQL).
14% and 4.3%: Nvidia's biggest drop on Thursday and its loss at the close (WSJ).

What happened

Business moved

Jensen Huang resigned as Nvidia's CEO and left its board, effective immediately. CFO Colette Kress is interim CEO, and director Harvey Jones becomes Nvidia's first chair (Nvidia, SEC 8-K).
Nvidia shares fell as much as 14% on Thursday and closed down 4.3%. Earnings are still due on 17 November (WSJ, Nvidia).
An internal Nvidia post asking the board to bring Huang back has passed 1,600 comments (Wired).
Anthropic expects to price its IPO on 18 November and start trading on 19 November, and reported $17.0B of third-quarter revenue (Anthropic).

Models & science

Google released Gemini 4 Hydrogen, now behind Google Search and the Gemini app, and an Orchestrator in Antigravity that pairs it with Carbon (Google).
Google DeepMind released AlphaGenome 2, trained on more than 100 million cell profiles across 68 tissues, for research use (Google DeepMind).
Runway released Praxis-1, an open-weight robotics model (Runway).
IBM linked two of its Nighthawk quantum processors at an average 1% error rate, a step toward its Starling machine due in 2029 (IBM).

Safety & policy

PostgreSQL fixed 52 vulnerabilities, including a 22-year-old code-execution bug. Upgrade to 18.7, 17.12, 16.16, 15.20 or 14.25 (PostgreSQL).
An independent analysis found Nvidia Blackwell fingerprints in six recent open models from DeepSeek, Moonshot, MiniMax and Ant Group (Will DePue, Unemployment Labs).
Two OpenAI engineers proposed a draft standard for agents to pass a user's country, masked IP range, time zone and locale to websites. It has not been adopted (IETF HTTP working group mailing list).

Platforms shifted

Brave's Search API began accepting per-search payment from agents through x402 (Brave).
Microsoft announced Microsoft Classroom, in preview early in 2027 and at no extra cost on current A1, A3, A5 and A7 education plans (Microsoft).
ElevenLabs released Scribe v3 and Scribe v3 Live for speech-to-text (ElevenLabs).
Rust 1.100 stabilised the never type and added dependency cooldowns to Cargo (Rust project).
Microsoft 365 had a service outage on Thursday afternoon, US time. Microsoft blamed servers that went offline and overloaded the rest (Microsoft).

We left out some launches that are still waitlist-only, a few tools that shipped weeks ago, and a run of Hacker News stories with nothing in them for founders.