Agents now create half of Notion's pages and booked four Uber rides in Halloween week. They take the work people hand off, not the tasks that take 50 seconds.

Tools

Build & ship

• ⭐ Parcel 3 alpha — A ground-up rewrite of the Parcel bundler in Rust, up to 100× faster than v2 on the same compilers, with native plugins in Rust, Go, Swift or C. Install parcel@next to try it. TypeScript .d.ts generation is not back yet. · Open source (MIT)
• Firenook — A drop-in replacement for the Firebase emulators that reads your existing firebase.json. It claims 40× faster median writes and 10× less peak memory, with no Java to install, and keeps your data across restarts. · Open source (Apache-2.0)
• faster-chrome-devtools-skill — An agent skill plus a dependency-free Node CLI that drive Chrome over the DevTools Protocol directly, with no browser MCP server. Screenshots go to files, so the agent loads them only when it needs them. · Open source (MIT)
• Graphify — Turns a codebase and its docs, SQL schemas, configs and PDFs into a knowledge graph you can query. Code is parsed locally with tree-sitter, with no LLM involved. Docs and media use your assistant's model. · Open source (Apache-2.0 / MIT)
• Testifly — Autonomous QA: connect your web app and it discovers user flows and generates tests for them with no test authoring. · Free trial
• Rivet — A YC W23 orchestrator for agentic workloads, built in Rust on durable actors. · Open source (Apache-2.0)

AI & agents

• ⭐ Images 2.6 — OpenAI's image model, 50% faster, with GIF creation and more precise editing. In the API, GPT-Image-2.6 Flare is built for speed and Sunburst for detail. · In ChatGPT and Codex on all tiers · API priced separately
• ⭐ Ripple-65B Preview — Ricursive's dense open-weight model with up to 1M tokens of context. It reports 80.52 on TerminalBench 2.1, ahead of much larger US open-weight models, for $32M in total R&D. Benchmarks are the company's own. · Open weights (Hugging Face, Together AI, OpenRouter and others)
• Kimi K3.1 Pro — Moonshot's 2.8T-parameter natively multimodal agentic model, downloadable from Hugging Face. · Open weights (Kimi K3.1 Pro License)
• SWE-2 Flash — Cognition's small coding model, trained with a mid-training stage before reinforcement learning. It runs at 3× the speed of SWE-2. · In Devin Desktop and CLI, at one-fifth of SWE-2's price
• AgentKit — Scalekit's gateway that connects your agent to 600+ apps, with per-user tokens, scoped tools and a log of every call. · Free to start
• MachineMail — Gives your agents their own email inboxes and attaches a trust verdict to every incoming message. Very early. · Open source (MIT)
• Stroq — Scans your Claude Code, Codex and Cursor sessions for known secrets, then adds local hooks that trace risky commands back to the content that prompted them. No telemetry. · Open source (Apache-2.0)

Design & create

• ⭐ Create UI — A Figma design system and a React component library that stay in sync: 80+ components and 73 ready-made blocks, plus an MCP server and skill so your agent builds with the same parts. · Free code library · Figma kit is a one-time payment
• Drawesome — A drawing toolbar for React: seven pens, an eraser, and SVG and PNG export, with no dependencies beyond React. · Open source (MIT)
• Thinking Orbs — Animated "thinking" and loading indicators for AI interfaces, in nine states, for React, React Native and SwiftUI. · Open source (MIT)
• Arsaze — A full video editor for creators that cuts silences and filler words in one pass and adds captions timed to every word. · Free to start
• Pangram Image 1.0 — Detects AI in images and video and shows a heatmap of where it appears, even after edits and compression. Comes with a Chrome extension and an API. · Free to try

Resources

Steal the template

• AlphaEvolve: Google DeepMind's minimal implementation of its Gemini-powered algorithm-discovery agent, plus the code and data behind the results it has found.
• The Stacks Project, formalized in Lean: Axiom's machine-checked library of the whole Stacks Project, the open reference for algebraic geometry. Researchers and agents can import it directly into new proofs.
• Tokenizer Mono: George Mandis's typeface built on IBM Plex Mono that shows LLM token boundaries as you type, using only OpenType features. It has 28 styles, one for each tokenizer.

Learn the craft

• MCP Explained: an interactive guide to the Model Context Protocol that shows every JSON-RPC message on the wire, step by step.
• Enterprise Managed Auth for connectors: Anthropic's docs on letting an organisation's users auto-connect to your MCP server through their identity provider, with no OAuth consent screen. Claude-issued assertions are in beta.
• How we built Alloy: Firecrawl on the internal workspace that gives agents its brand rules and past approved work. It now has 63 reusable animation packages and editable product demos.

Benchmarks you can re-run

• Village LLM Atlas: maps the best model at every price point. Check where your current model sits before your next renewal.
• DOOM's DNA: compares the original DOOM source with 43 descendants. In a blind test, functions in a Go translation were matched to the right game subsystem 60% of the time, against 20% for shuffled labels.

Reads

☕ Under 5 minutes

• Retire the word "context": Karri Saarinen of Linear argues that more stored context does not tell an agent what the work is for. Give it a short brief that explains why decisions were made.
• The Next Era of Notion: Ivan Zhao on why Notion is turning into the place other companies' agents keep their context.
• The Financial Engineering Essential for AI in 2027: Tomasz Tunguz on GPU costs, burst compute and why your AI contracts are about to get longer.
• Bad AI doesn't smell: Clearleft on why wrong AI research analysis looks as confident as right analysis. Do your own synthesis first and use the AI only as a check.
• Some reasons SaaSpocalypse might not be as bad as we thought: AI lowers the barrier to building software. It does not remove it.

🍵 5–10 minutes

• The top 1% of AI spenders stalled in Q3: SaaStr goes through Ramp's AI Index and Anthropic's amended S-1 together.
• Your customers don't just want an agent. They want agents.: Irrational Labs' 1,014-buyer test of agent labels on the same product page.
• Copying Cope: SuperMe's CEO on four startups copying material parts of its product in a couple of months, and why distribution now matters more than the product.
• Agents struggle with the same bad UX we do: computer-use agents fail at the same confusing screens that trip up people, which makes them a cheap way to find your design debt.
• Designing product demos people can actually follow: a demo is a sequence of state changes. Design how attention moves from one to the next.
• Why 10,000 experiments work better than 10,000 hours: break a career goal into microskills and run small weekly experiments on each.
• You'll need nerves of steel to work with AI agents: over 100 hours with Cursor's Composer 2.5, including unrelated code deleted and tests that skipped the API routes.
• Agentic technical debt: Kai Waehner on where retries, approvals and audit trails belong once an agent works across more than one system.
• How we made clarity.js 4.5% smaller: Microsoft ran more than 50 automated experiments to cut 1,257 gzip bytes. The biggest saving could not ship because it would break Angular sites.

📚 Longer, worth it

• A thousand fewer paths to root: OpenAI's full write-up of its pipeline, from candidate findings through validation, exploits and patches, with the cost of each stage.
• Everyone is building AI agents. Who's designing the stop button?: let AI interpret what the user wants, and use deterministic steps, confirmations and recovery for anything consequential.
• The end of feature discovery as we know it: Elena Verna on what growth teams do once the software picks which feature to use.
• Critical thinking when agents do the work: Addy Osmani on the "orchestration tax": when you produce more than you can stand behind, you are no longer reviewing, you are hoping.
• Nobody here writes code by hand anymore: how Intento's team moved to agents, and why its early failures came from context that lived in people's heads.
• AI companies should own their metering, not the entire billing stack: build usage measurement, pricing logic and balances yourself, and buy invoicing and revenue recognition.
• Scaling Slack and GitHub with Maude Lemaire: Gergely Orosz interviews the engineer who wrote Refactoring at Scale about load-testing and migrations at two of the largest developer platforms.
• Dharma Field Extractor: how a 2B-parameter self-hosted model beat frontier APIs at pulling structured fields out of documents, by the team's own benchmark.
• Silicon Valley's Very Bad Night: Matteo Wong on an election night when voters turned against data centers.

Numbers

• 948: working Linux privilege-escalation exploits produced in one automated run, at $1,279 each in API credits (OpenAI).
• $5.40 billion: revenue Anthropic added in the third quarter, down from $6.84 billion in the second (Anthropic S-1/A; SaaStr).
• 82%: rise in Anthropic's blended GPU-hour cost over two quarters, from $3.10 to $5.65 (Tomasz Tunguz, from the S-1/A).
• $9,568: monthly AI spend per employee at the top 1% of firms on Ramp in September, down from $9,873 in July (Ramp AI Index; SaaStr).
• 4: Uber rides ordered by AI agents in Halloween week, alongside about 500 Uber Eats orders, out of roughly 320 million trips and orders (Uber; TechCrunch).
• ~50%: share of pages in Notion now created by agents. Notion's MCP volume is up about 60× this year (Ivan Zhao).
• 100,000: failed tasks that Zapier says self-repairing Zaps saved in their first week (Wade Foster).
• 3%: share of price points where an open-weight model is the best buy, down from 41% in June (Village).
• $32 million: Ricursive's full R&D cost for Ripple-65B Preview, which it says is a tenth of the previous open-weight leader in agentic coding (Ricursive).
• 6.3 points: rise in stated purchase interest from calling a product a team of agents rather than a single agent (Irrational Labs).
• 35%: drop in median time to first token for 500k to 1M-token requests using Anthropic's message threads (Anthropic).
• 1,271: roles affected in Canva's restructure, 22% of its workforce (The Age).

What happened

Models & science

• OpenAI released Images 2.6 in ChatGPT, Codex and the API, 50% faster and with GIF creation (OpenAI).
• Moonshot released open weights for Kimi K3.1 Pro, a 2.8T-parameter multimodal model (Moonshot AI).
• Ricursive released Ripple-65B Preview, an open-weight model that reports 80.52 on TerminalBench 2.1 (Ricursive).
• Thinking Machines launched Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active, on its Tinker platform with closed weights. It plans to release the weights under Apache 2.0 next week after external safety testing (Thinking Machines).
• Cognition released SWE-2 Flash, its small coding model, in Devin Desktop and CLI (Cognition).
• Tencent announced Hy4, a model with visual understanding, and said weights and system cards will follow in the coming days (Tencent).
• Baidu released ERNIE 5.5 Flash Base, open base weights for a 280B-parameter mixture-of-experts model with 14B active, pretrained on 33.5T tokens (Baidu).
• Axiom formalized the entire Stacks Project in Lean, which it calls the largest integrated library of formalized mathematics yet (Axiom).
• Google DeepMind published a minimal open implementation of AlphaEvolve (Google DeepMind).

Safety & policy

• OpenAI's security pipeline produced 948 Linux privilege-escalation exploits and candidate patches, and it used them to prioritise reports to maintainers (OpenAI).
• PHP2Shell: an AI-assisted researcher turned a PHP out-of-memory bug into remote code execution, and the fix took two months (IEEE Spectrum).
• A Claude Code user opened an issue saying the tool sent their email address to third-party sites in its User-Agent header during web searches (GitHub).
• Voters in at least 10 states weighed dozens of proposals to restrict data-center development, and several towns recalled mayors over data-center deals (The Atlantic).

Business moved

• Anthropic's amended S-1 shows third-quarter revenue of $16.96 billion. Tomasz Tunguz reads it as a $567 million operating loss, after a $974 million profit in the second quarter (Anthropic S-1/A; Tomasz Tunguz).
• OpenAI agreed to acquire Pierre Computer Company, the maker of Code Storage, to expand Codex's tools for building and reviewing code (OpenAI).
• Meta acquired Conductor, which runs several coding agents in parallel. Conductor Cloud goes read-only on 13 November and shuts down on 20 November (Conductor).
• Canva is restructuring 1,271 roles after its growth slowed for the first time in a decade. Departing staff get at least six months' salary (The Age).

Platforms shifted

• Anthropic put message threads into public beta, so long-running agents send only new messages each turn (Anthropic).
• Claude added Enterprise Managed Auth, letting organisations connect users to MCP servers through their identity provider (Anthropic).
• Zapier's Next Gen Zaps now read the error when a step fails, work out and test a fix, and publish a new version of the Zap (Wade Foster).
• Cloudflare launched analytics and per-agent access policies for personal AI agents visiting websites (Cloudflare).
• Perplexity's Decisions API can now search the web before it answers (Perplexity).
• Go 1.27 made encoding/json/v2 generally available, rejecting invalid UTF-8 and duplicate names by default (Go).
• Safari in the iOS 27.3 and macOS 27.3 betas adds vertical tabs and an Adaptive Dark Mode for sites that lack one (9to5Mac).
• Adobe's Premiere mobile app is adding licensed Universal and Warner tracks, limited to 60 seconds per video and personal use (Digital Trends).

We left out a run of acquisitions and hires announced only in social posts, a batch of meal-planner app growth claims we could not trace to a source, and the usual sponsor placements.