Replit's agent can now buy your supplies and run your Google Ads. This week, agents got wallets, and seat pricing got weaker.

Today's haul: 16 tools · 5 resources · 22 reads · 12 numbers · 18 things that happened. Every tool, resource and read below has a working link. No link, no listing.

The 60-second catch-up

Agents can spend money now. The limit has to sit at the bank, not in the prompt.

Replit Agent can shop merchant sites for one-time business supplies and stop at checkout for the owner to approve the total. It can also build and launch Google Ads campaigns (Replit). Meta's Muse now phones US businesses for paid subscribers (TechCrunch). The ad industry's trade body published a framework that separates purchases a human made, ones AI assisted and ones AI executed (IAB). Alan Williamson's advice is blunt: don't trust a provider's budget cap. Put a hard limit on a virtual card or prepaid credits, and add guards in your own code (Alan Williamson).

→ do this: give every agent and API key its own virtual card with a monthly limit you can afford to lose.

One founder can run dozens of agents. They still need one boss.

Nathan Broadbent runs DocSpring with 49 Claude Code sessions, 26 of them active, each on one job: one chases Sentry errors, another works through small bugs (DocSpring). Anthropic researchers tested the other way to organise agents, one per user with no one in charge. Across 25 scenarios, Claude Opus 5 teams reached 30% of the best outcome without a coordinator and 64% with one. The agents mostly acted "too little and too late" (Anthropic).

→ do this: if you run more than three agents, give one of them the job of assigning work and checking it. Don't let them negotiate with each other.

AI is shrinking seats, so price on the job, not the head.

Alex Turnbull closed signups at Groove, his profitable support software company, at $4.7M ARR. Customers were downgrading seats as AI shrank their support teams, more often than they cancelled (Alex Turnbull). Jason Lemkin's case is the flip side: an agent that does work a person used to do gets paid from the labour budget and can double contract value. Most AI features don't move it at all (SaaStr). Payments are a third route, because they grow with transactions, not headcount (Euclid Ventures).

→ do this: count how many of your customers have fewer seats than a year ago. If that number is rising, test a price tied to the work done.

Tools

Build & ship

ggml v1.0.0 — The tensor library under llama.cpp reaches 1.0. Old APIs are gone, FP8 is in, and it's faster across CUDA, Metal, Vulkan and WebGPU. · Open source (MIT)
Qwen Code 1.0 Preview — Alibaba's open coding agent now uses the agent harness from its Qoder IDE and installs natively on macOS, Linux and Windows, with no Node.js. Preview, so interfaces may change. · Open source (Apache-2.0)
UI SDK — Vercel Labs' library of 50+ interface behaviours (focus, typeahead, popups, drag to reorder) that coding agents combine into components for React, Vue, Svelte, Solid and Web Components. Beta. · Open source (Apache-2.0)
Relax — An experimental Rust JIT from Makepad. It runs a Cargo package from source and patches your edits into the running program without losing its state. For development builds only. · Open source (MIT)
GraphQL over HTTP spec — The first official spec for GraphQL over HTTP. A new media type lets errors use real 4xx and 5xx codes instead of a 200 with an error inside. · Free (open spec)
Clayzo — An animation engine for coding agents. Claude Code, Codex or Cursor writes the motion as code, checks each frame and ships it to the web or to Swift. · Free

AI & agents

⭐ Replit Agent, commerce and ads — The agent buys one-time business supplies and waits for you to approve the total, paid through Link and Visa. It also builds and launches Google Ads campaigns. · Paid · buying on Core and Pro (US) · ads on Pro (worldwide)
⭐ Perplexity Decider v1.3 — Classifier models that return probabilities for yes/no, multiple-choice and scoring questions in one pass. New in v1.3: "locate", which pulls exact spans such as names or personal data out of the text. A Flash version is smaller. · Open weights (Apache-2.0)
Claude Meetings — Claude joins a Google Meet as a visible participant. It takes notes, answers questions in the chat and builds docs or slides during the call. Beta, rolling out now. · Team and Enterprise plans
Atlas — World Labs' model that works across text, images, video and 3D to generate and rebuild scenes, now open to everyone. Frameboy lets you walk around its worlds in real time. · Paid (credits) · no price stated
LiquidCodec — Liquid AI's small (~74M-parameter) streaming audio codec, the one inside its LFM3 audio models. Built for voice apps that run locally. · Open weights · Liquid LFM licence
ChatGPT Dots, custom instructions — ChatGPT's personal agents now follow your rules, such as "show me a draft before sending any emails." · No price stated
Muse calls businesses — Meta's Muse agent phones US businesses for you: the cable company, the DMV, a restaurant with no online booking. · Paid plans from $20/mo

Design & create

Iconden — A community gallery of iPhone app icons. Browse, track redesigns and submit icons you spot. · No price stated
Isometric SF — The whole Bay Area as a zoomable isometric pixel-art map, each tile painted by an AI model from Google Maps 3D captures. · Free · tiles CC BY-NC 4.0

Growth & ops

Brave 1.98, photo ID stripping — Brave on desktop now strips Meta's hidden "FBMD" identifier from Facebook and Instagram photos when you download or upload them. · Free

Resources

Steal the template

How we booked 321 calls on LinkedIn in two months: Needle's play is to message people right after they show a buying signal, with a person approving every note. It includes reply rates for each signal.
How to prevent runaway AI bills: Alan Williamson's setup of virtual cards, prepaid credits and code-level guards, so a loop can't spend more than you planned.
IAB's AI-era attribution blueprint: A shared way to label a sale as human-owned, AI-assisted or AI-executed, and decide which channel gets the credit.

Benchmarks you can re-run

Agent Anomaly Detection on benchmarks Google didn't write: Google tested its agent-security monitor on public benchmarks. It caught most hijacks that ran, and only 4 of 18 poisoned-memory cases.
Stanford SCALE: the evidence on AI in K-12: 52 causal studies, up from 20 in March. Letting AI do the work hurt how students did later without it.

Reads

☕ Under 5 minutes

Setting the pace: Sequoia's Alfred Lin on why early speed predicts success, and how to stop process from slowing you down.
One whale or six smaller customers?: The maths on a discounted $3M customer against six full-price ones, once you count gross margin.
Design should make it easier to ship: Karl Koch on judging design by how easy it makes shipping, not only by how it looks.
Cost per task, not price per token: Simon Willison on why per-token prices mislead once reasoning models use very different numbers of tokens.
Put MCP next to your API, not on top of it: Jerome Louvel on serving REST and MCP side by side so you don't maintain business logic twice. A vendor post.

🍵 5–10 minutes

A great B2B agent doubles your ACV: Jason Lemkin on why agents get paid from labour budgets, and AI features mostly don't.
Why I stopped selling a profitable SaaS: Alex Turnbull on closing Groove's signups and starting again as Helply.
Tactical team changes from AI: Stay SaaSy on tiny teams testing new markets, and who should run them.
Monetizing payments in vertical AI: Euclid Ventures with Payabli's co-CEOs on earning from payments when seat counts shrink.
Rebundling your data stack is inevitable: PostHog's Lizzie Epton on choosing a data bundle without getting locked in. A vendor post.
What's defensible for small builders now: Flavio Copes on what you can still own when AI labs ship apps next to yours.
Everyone is learning the wrong lesson from Hugging Face: Karthik Tadepalli argues the agent break-out at OpenAI and Hugging Face was an alignment failure, not a security one.
Migrating our email API to Go: Resend rebuilt its core API in Go in seven weeks and matched the old behaviour exactly.
From Vue and jQuery to Stimulus: DNSimple moved about 50 Vue components back to server-rendered Rails. The code mostly moved rather than shrank.
Why a button takes a month: TJ Pitre converts IBM Carbon's button between React and Figma. 40 of 84 variants passed.
Confessions of a recovering full-time employee: Ibrahim Bashir on why he kept going back to full-time jobs, and the portfolio career he chose instead.

📚 Longer, worth it

Operating a SaaS when AI agents write all the code: Nathan Broadbent on running DocSpring with dozens of Claude Code sessions, and what still needs a human.
Worse together: Anthropic researchers on why agents that each serve one user do far worse than agents with a coordinator.
Yes, you still need a technical co-founder: Gal Ratner on how a vibe-coded prototype lets you avoid finding paying customers.
Orienting to one-year AGI timelines: Nikola Jurkovic on what non-technical people should do if superintelligence arrives by early 2028.
Eliminating nondeterminism in LLM inference: Nathan Barry runs an LLM on integer maths only, so output is bit-identical on any hardware. It's slower.
After 600 experiments with JEPA world models: Haddy Alchaer found that how you plan with the model matters more than how well it predicts.

Numbers

49: Claude Code sessions Nathan Broadbent manages for DocSpring, 26 of them active (DocSpring).
30% vs 64%: share of the best outcome Claude Opus 5 agent teams reached without and with a coordinator (Anthropic).
$4.7M: Groove's ARR when Alex Turnbull closed signups in March (Helply).
321: calls Needle booked on LinkedIn in two months (Needle).
34.2% vs 9.9%: Needle's reply rate from people who commented on a post about its category, against followers of its own company page (Needle).
58 → 31 ms: Resend's median email-send latency after moving its API to Go. Memory per task fell from about 1.3 GB to 180 MB (Resend).
87%: share of executed goal hijacks Google's agent monitor flagged on outside benchmarks (Google).
52: causal studies of AI in schools in Stanford's review, up from 20 in March (Stanford SCALE Initiative).
$230.9M: Cerebras's core third-quarter revenue, up 70% (Cerebras).
$74B+: SpaceX's planned AI computing sites in Texas and Arizona (SpaceX).
$400M+: crypto North Korea's "FaceStealer" scheme stole in 18 months, US prosecutors say (US Justice Department).
~$947M: Nvidia shares director Mark Stevens sold in September, with no trading plan ticked on the filings (SEC Form 4s).

What happened

Business moved

The Justice Department has several probes open: one antitrust probe of Nvidia, and two into Megaspeed and Supermicro. Officials say there is no sign of charges against Nvidia or Jensen Huang (Semafor).
Huang agreed to testify before Congress, and Senator Tim Scott asked him to appear before the Senate Banking Committee (TechCrunch).
Nvidia's board approached Arm CEO Rene Haas about the top job. The board says it has not formally engaged any candidate (Financial Times, Reuters).
In August, CFO Colette Kress pushed back on Huang's funding of startups that then bought Nvidia GPUs (Wall Street Journal).
Anthropic's IPO was about six times oversubscribed (Bloomberg). In its roadshow, co-founder Tom Brown said Claude ported its prompt processing to AMD chips in about three days (Anthropic).
Cerebras raised its full-year outlook after core revenue rose 70% (Cerebras).
SpaceX announced more than $74B for AI computing sites: about $50B in Cochise County, Arizona, and more than $24B at Starbase, Texas (SpaceX).

Safety & policy

Four nonprofits sued Anthropic, Google, Meta, OpenAI and SpaceXSI in San Francisco Superior Court to freeze new frontier models (AP).
Three OpenAI national-security staff spent 17 minutes on a video call with Russian pranksters posing as Ukraine's spy chief. OpenAI says it investigated and took steps to prevent a repeat (The Intercept).
OpenAI told lawmakers it will back bills restricting China's access to AI chips and chipmaking equipment (Bloomberg).
The European Commission proposed a Digital Fairness Act banning dark patterns and hard-to-cancel subscriptions, and limiting loot boxes for minors. A companion rule would allow fines of up to 6% of worldwide turnover (European Commission).
US prosecutors disrupted North Korea's "FaceStealer" scheme, which used fake video calls with recordings of people victims knew, and separately seized more than $1.2B stolen from Binance (US Justice Department, FBI).
Citizen Lab found security agencies in many countries using Penlink's Webloc, which tracks people through location data from phone ads (Citizen Lab).

Platforms shifted

The FAA stopped departing flights nationwide for 4 hours 21 minutes on Thursday. Transportation Secretary Sean Duffy blamed old air-traffic routers overloaded with data, and more than 7,000 flights were cancelled (NBC News).
Anthropic opened a private beta of self-hosted Claude Managed Agents, which run sandboxes, credentials and session history inside a company's own cloud (Anthropic).
Google opened preorders for Gemini glasses from Warby Parker (from $395) and Gentle Monster (from $449), in the US and South Korea, in stores 10 December (Google).
World Labs opened its Atlas model to everyone (World Labs).
ggml, the library under llama.cpp, reached version 1.0 (ggml).

We left out a few tools that launched more than a week ago, some launches still behind a waitlist, and Hacker News stories with nothing in them for founders.