Agents can now publish your app, and AWS will box them in for free. The hard part has moved to checking what they did.

Tools

Build & ship

• ⭐ Spacefast — Automattic's publishing layer for your agent. Connect it to Claude, ChatGPT, Cursor or Codex and the agent ships an app with hosting, a custom domain and a SQL database. Every publish can be rolled back. · Free plan with $1 of usage a month · Go $5/mo
• ⭐ AWS Nitro Sandboxes — Runs each EKS pod, ECS task or lightweight VM in its own hardware-isolated VM on a shared EC2 instance. A one-line change, built for agents running code a model wrote. · No extra charge · you pay for EC2 as usual
• Modern Web Guidance 1.0 — An agent skill and CLI that feeds your coding agent current guides to web APIs, so it stops writing 2019-era JavaScript. Version 1.0 adds a scan for legacy patterns and is now run jointly by Chrome, Edge, Firefox and MDN. · Open source (Apache-2.0)
• Reticle — A dev-only SDK for your web or desktop app that lets your coding agent drive the running app, read its network, state and console, and say whether "done" is true, with the file and line to fix. · Open source (Apache-2.0) · hosted Harness on credits
• Skullet — Skeleton loaders for React, Vue, Svelte and plain JS that measure your real component with sample data, so nothing jumps when the data arrives. · Open source (MIT)

AI & agents

• ⭐ PostHog Self-Driving — Agents that read your PostHog errors and analytics, write a ranked inbox of what to fix and, if you turn it on, open the GitHub pull request. Now out of beta. · Per run, in AI credits at $0.01 each · first month includes $10
• recalc — A desktop spreadsheet where a cell can hold a formula, a terminal command or a live Claude Code or Codex agent. Test cells re-run and turn red when an agent's "done" is wrong. Early beta. · Open source (Apache-2.0)
• Voyager — A Finder replacement for the Mac with agent chat built in. Select a folder or files as context and point Claude Code, Codex or Hermes at them. · Free · paid plans not yet available
• airgap-check — A CLI that finds what will break when your AI app runs offline: first-use model downloads, silent telemetry, secret names in .env. It can run your smoke test with the network cut and log every attempt to call out. · Open source (MIT)
• ElevenAgents Campaign Manager — Give your ElevenLabs voice agent a contact list and a goal and it works through the list by phone or WhatsApp, with retries, attempt caps and do-not-call lists. · No price stated · part of ElevenAgents

Design & create

• Rainbrand — Stores your brand guide, logos, fonts and voice so agents can fetch them over MCP and check finished work against them. Formerly Bloom. · From $20/mo

Growth & ops

• Orphex MCP — One MCP connection from your ad, analytics and commerce accounts, such as Google Ads, Meta, Shopify and RevenueCat, to your AI assistant. Ask about performance, then approve the changes it proposes. · Free tier · Read $39.99/mo · Action $149.99/mo
• Databox CLI — For Databox users: a command line and agent skills covering the Databox API, so Claude Code, Cursor or Codex can create datasets, push data and build metrics. · Open source (MIT) · needs a Databox account

Resources

Steal the template

• The 2027 funnel audit: Sam Kuehnle's worked audit for annual planning. His example shows how a $30 cost per demo turns into a very expensive opportunity once you follow the conversion rates down.
• The secret behind viral product launches: Kyle Jeong, who runs launches at Browserbase, on what they cost. A launch video easily runs $100k–250k, and 1M guaranteed views through an agency is $35k–40k up front.

Learn the craft

• Safely scaling multi-agent teams: Anthropic's write-up of running 1,000 agents on one proof, with the sandbox, command classifier and monitor that kept it contained. A shared machine broke at about 100 agents.
• Not everything needs an LLM: Jina Yoon at PostHog puts a small decision model in front of the LLM. 98.7% of summary requests now skip the LLM call entirely.

Benchmarks you can re-run

• TIS-100 Bench: 16 models on 55 puzzles from a programming game, where a solution counts only if it passes 100,000 tests. The models found 133 new best solutions on 18 puzzles.

Reads

☕ Under 5 minutes

• When does the US become a net importer of tokens?: Tomasz Tunguz's model puts the switch between 2028 and 2030, depending on how many planned data centers never connect.
• The thing I talk to most: Shane Mac on the internal agent he built three weeks ago, which now proposes product decisions and scores the company every week.
• The new bottleneck is intent: Zachary Kiihne says fixing bugs got cheap, so the scarce work is defining what the product should do.
• Why is AI slop so frustrating to read?: Zonder Labs builds a scale for unclear sentences and runs agent replies against Stack Overflow answers.
• The race to build a "Find My" for AI chips: IEEE Spectrum on tracking GPUs after a man was charged with shipping over $300 million of GPU servers to China.
• The right decision: Sean Byrnes argues there is a right answer even when you cannot know it, so commit and move on to the next call.
• AI usage generates a fragile codebase: Evan X. Merz, a former Meta engineer, on what happens to a codebase when nearly all of it is generated.
• After 7 months, Glasswing overwhelms maintainers: Patrick Garrity counts what Anthropic's bug-hunting project found, and how little of it got fixed.

🍵 5–10 minutes

• State of the tech job market: late 2026: Lenny Rachitsky's data on who is hiring: PM openings up, engineering flat, design at a record low. Paid after the first six takeaways.
• How Faire developed product builders: Faire CTO Thuan Q. Pham on turning 214 engineers into people who ship features end to end, with the before-and-after numbers.
• The great software IP writedown: Martin Alderson argues software margins came from a shortage of builders, and agents remove the shortage.
• How Product Hunt almost died: Ryan Hoover on pitching more than 50 Series B leads and hearing no from all of them, and what he would do differently.
• Own your means of production: Retool's David Hsu reads Thrive Holdings' $15.3 billion deal for ADT as the start of buying companies and running them with AI.
• What does a design engineer do?: Nurkhon Akhmedov's team process: no Figma, one branch, a three-to-four-day sprint, then a hand-off to engineering.
• How AI helped build the new Dropbox: Pravi Garg on a test clean-up that was put off for two quarters, split into 53 tasks and finished in a week.
• Claude designs electronics from start to finish: Anthropic had Fable 5.2 design three devices, from a speakerphone to a virtual pet. None has had electrical validation or certification.
• Patent claims stop work on DisplayLink drivers: Joe Brockmeier on a Linux developer who improved a driver with LLM help, and the patent letter that ended the work.
• Will Anthropic grow into its compute bill?: Ornn works through Anthropic's S-1 and concludes it would need to be roughly a $200 billion-a-year business to carry its commitments.

📚 Longer, worth it

• Should you put AI in your positioning?: FletchPMM's long look at when to lead with AI and when to leave it out, with AI CRMs as the cases.
• Don't make a human do what AI is amazing at: John Nunemaker on a weekend of 113 merged pull requests and the 107 review rules he wrote along the way.
• Software correctness and supply-chain security: Gergely Orosz interviews Sean Allen of Antithesis on making illegal states impossible to express, and why there is no simple fix for supply-chain attacks. Paid.

Numbers

• ~113,000: US software engineers hired from May 2025 to April 2026, less than half the 2022 peak of 239,000 (Live Data Technologies).
• 4.6×: one-year rise in open forward-deployed engineer roles, now about 1,500 (TrueUp).
• 3.8: open PM roles for every open product design role, up from 1.9 in early 2023 (TrueUp).
• $655K: median annual recurring revenue per employee at top-growth software companies above $100M ARR (ICONIQ).
• ~3×: rise in the typical Faire engineer's complexity-weighted pull-request output, across 214 engineers (Faire).
• 9.8% vs 6.4%: rise in total family health premiums against the rise in gross pay for the same US employees, 2025 to 2026 (Rippling).
• 51%: weekly use of standalone AI tools across six countries, up from 34% a year earlier (Reuters Institute).
• 217 of 37,966: findings from Anthropic's Glasswing project that have been fixed (VulnCheck).
• $895B: Anthropic's committed cloud and data-center spending, as tallied from its amended S-1 (Ornn, DCD).
• 35 seconds: how fast OpenAI's Control Room halts research model output after a stop decision (OpenAI).
• 28.7 hours: how long 1,000 Fable 5.2 agents took to reproduce a formal proof of Fermat's Last Theorem, against 11 days in September (Anthropic).

What happened

Models & science

• Google DeepMind introduced AlphaProof Code, which used Gemini agents to write more than 6 million lines of Lean proofs about Signal's protocol code in four weeks (Google DeepMind).
• NVIDIA showed GR00T 2, a robot model that succeeds on 40% of unseen tasks with no examples and 95% after five demos and a 30-minute fine-tune. Code is due in December (NVIDIA).
• Anthropic says Fable 5.2 designed three working electronic devices, none yet electrically validated or certified (Anthropic).

Safety & policy

• OpenAI described Control Room, the system that registers every research inference job and can halt model output across its infrastructure (OpenAI).
• Fireworks AI says attackers linked to an "LLMJacking" group reached its development machines through a third-party cloud provider, and accessed billing data and, for some customers, partial model weights and datasets (Fireworks AI).
• The US now requires federally funded researchers to buy synthetic DNA from approved suppliers that check customers and screen orders, phased in over 340 days (Science, White House OSTP).
• Meta expanded AI age estimation worldwide to move teens who lied about their age into Instagram Teen Accounts (Meta).
• Cisco is adding Claude Mythos to the red-teaming in its AI Defense product (Cisco).

Business moved

• SpaceX plans a gas power plant and data center in Texas whose first phases cost about $24 billion, on land carved out of Brownsville (Bloomberg, Texas Tribune).
• Val Town stopped new signups. It goes read-only on 10 December, shuts down on 10 January and will be open-sourced on 1 February (Val Town).
• beehiiv removed all internal job titles, including CEO (Tyler Denk).
• Adobe and Disney agreed a multi-year alliance to build custom image models trained on Disney's own characters and styles. No terms were disclosed (Adobe).
• ChatGPT Enterprise is now on AWS Marketplace, so it can count toward AWS spending commitments (OpenAI).
• Automattic launched Spacefast, a publishing platform for agents (Automattic).

Platforms shifted

• Bloomberg's Mark Gurman reports Apple's new CEO is pushing to turn Siri into an autonomous agent as soon as next year (Bloomberg).
• Logic Pro now sells a $49.99 Intelligence Usage Pack that extends its cloud AI features past the usage limits (The Verge).
• Microsoft released .NET 11, with C# 15 union types and built-in agent tooling (Microsoft).
• PostHog took Self-Driving out of beta and moved it from $15 per pull request to per-run pricing (PostHog).
• ElevenLabs released Campaign Manager for ElevenAgents (ElevenLabs).
• Modern Web Guidance reached 1.0 and is now maintained jointly by Chrome, Edge, Firefox and MDN (Google Chrome).

We left out a sponsor's competitor-tracking tool, the model launches we covered yesterday, and some Hacker News curiosities.