Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · November 11, 2026 · afternoon edition

The builder launches of November 11 give agents a place to check their own work rather than a smarter brain, led by AWS shipping a local copy of AWS for coding agents and Google pricing a yes-or-no judge at a dime per million input tokens.

AWS EmulatorGemini Decisions APIClaude CodeAnthropicPomeradoagent-infrastructureclaude-codemcp-securityagent-harnessfrontier-models

Trending AI Briefing: Wednesday, November 11, 2026 (afternoon ET)

Two of the biggest names in cloud spent Wednesday shipping things that make agents more checkable, not more clever. AWS released AWS Emulator, a free local copy of the AWS APIs built so coding agents can test their code without touching a real account, and Google launched the Gemini Decisions API, a model that only answers fixed questions and costs nothing for output. Both are bets that the bottleneck in agent work has moved from writing the code to proving it is right, and the same week's security fixes and contribution bans say the proving part is still mostly unbuilt.

What's hottest in AI news right now

AWS released AWS Emulator on November 11, a local stand-in for AWS that its own agents wrote. The launch ran as a guest post by Swami Sivasubramanian on Werner Vogels' blog, which says the emulator was built from more than 400 API models covering more than 200 AWS services and "runs free and offline on your own machine." A companion Amazon post by Matthew Goldberg says autonomous agents in Kiro wrote all of the code, starting with a working SQS prototype in about 15 minutes, and that "tens of thousands of fidelity tests" written from AWS's public API models and checked against real AWS decide whether the copy behaves like the original. The README says it needs no AWS account, ships as RPM and Debian packages on Linux and a container on Windows and macOS (a source build serves HTTPS on 127.0.0.1:9297), and installs as an agent skill with aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1 (AWS CLI 2.35.0 or later). AWS's News Blog calls it generally available. The catch: telemetry is on by default, AWS says it does not evaluate IAM or resource-based policies the way AWS does, the README says not to use it for production or security testing, the repo takes no pull requests, and while the README says per-service fidelity scores ship with each release, the November 8 releases said scores were "not available" and the latest, v1.0.322, shows none. It reached the Hacker News front page within hours. All Things Distributed · About Amazon · README · HN

Google launched the Gemini Decisions API on November 11, a separate endpoint that answers closed questions instead of writing text. It runs a new model, gemini-decision-1, that takes one shared input (text, images, audio, video or PDFs, up to 256K tokens) and up to 1,000 yes/no, multiple-choice or rubric-score questions per request, each answer carrying a confidence score from 0 to 1. Google prices it at $0.10 per million input tokens with output billed at $0, bills the shared input once per request, and quotes 150 to 250 ms end to end. The honest catch is that every speed and cost number is Google's own, including "nearly 5x cheaper" than Gemini 3.5 Flash-Lite on Google's benchmark suite, and that free-tier content may be used to improve Google's products, which matters when the input is an agent transcript. OpenAI's Decisions API, now generally available on gpt-6-luna, has the same shape, so decision endpoints are now a two-vendor category. Google blog · Decisions API docs

Claude Code 2.1.321 closed an MCP hole that could hand your refresh tokens to a different server. The release, published to npm on November 9, includes the line "Fixed MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them." Three more fixes in the same version touch the permission system: a read-only Bash command that ran outside the sandbox without a prompt, Bash permission checks that judged a file write from the wrong folder, and deny rules that missed the file a cp or mv into a directory writes. The same release made Claude Fable 5.2 the default Fable model. The changelog carries no dates, so the ship date comes from the npm registry; the current latest is 2.1.322. Anyone who connected Claude Code to a third-party MCP server over OAuth should update. CHANGELOG · npm

Anthropic launched the roadshow for its initial public offering on November 10. The notice says Anthropic, PBC has applied to list Class A shares on the Nasdaq Global Select Market under "ANT," offering 825,000,000 shares (801,351,354 from Anthropic, 23,648,646 from existing holders) at an expected $70 to $78 each, with a 30-day option for underwriters to buy 123,750,000 more. Goldman Sachs and Morgan Stanley are the named prospectus contacts. By simple arithmetic the base deal sits between about $58 billion and $64 billion at the range, a figure the notice itself does not state. For builders the practical question is whether public-market pressure changes pricing and model retirement schedules, and nothing in the filing notice answers it yet. Anthropic

A front-page Hacker News story about 48 AI agents "finding" a Maya pyramid in NASA lidar got corrected by its own author. Pietro Miotti (a Google software engineer, per the Hacker News headline) pointed 48 AI agents at about 1,200 square kilometers of NASA lidar, and in less than 72 hours went from idea to report to an official reply from Mexico's INAH saying no registered site sat at that location. A November 11 update says the site already appears in a 2025 paper by Estrada-Belli and colleagues, so the agents "rediscovered it rather than discovered it." The post names Gemini, ChatGPT, Claude and Grok as brainstorming help but not the model the 48 image-review agents ran on, the author says plainly that he is not an archaeologist, and INAH's field check is still pending. The lesson for anyone running search agents at scale is narrower than the headline: they are good at surfacing candidates and bad at knowing what has already been published. Pietro Miotti · HN

New tools and features worth actually trying

Atuin Threads. The shell-history tool now records Claude Code, Codex, Pi, Omp and OpenCode sessions and syncs them across machines, end-to-end encrypted, so you can atuin resume a session on your laptop that started on a server. Announced October 27 for Atuin 18.24 and on the Hacker News front page today. Honest tradeoff: capture and sync are opt-in for good reason, since a synced agent log holds whatever secrets you pasted into it, and team sharing is still being built. Atuin blog

Pomerado. A Y Combinator company (Launch HN today) that turns any website into an API or MCP server so an agent calls functions instead of clicking through pages. The free tier allows 100 tool calls a day. Honest tradeoff: its "2x faster, 10 to 1,000 times cheaper" claim comes from its own 30-task benchmark, it routes through residential proxies to get past anti-bot checks, and handing it site logins is a trust decision you should make on purpose. Pomerado

The AWS Emulator agent skill. If your agent writes code against S3, SQS or DynamoDB, installing the skill lets it set up the emulator and run its tests locally before anything reaches a real account. Honest tradeoff: turn telemetry off with EMULATOR_TELEMETRY=off, expect gaps the project tracks in its issues/things-that-do-not-work/ folder, and do not read a green local run as proof IAM policies behave the same in production. README

Klad. A painting and animation app where a coding agent draws through a CLI (klad stroke, klad fill, klad export) on the same canvas you use, with no image model inside, so every stroke stays editable data. Show HN today. Honest tradeoff: it is source-available under a PolyForm Internal Use license, free only for individuals and companies under 500,000 euros of revenue, with no installers yet and macOS as the only fully supported platform. Klad

Trending AI repos on GitHub today

Trendshift's daily board, read at about 15:15 ET on November 11, ranks repos by momentum score, not star totals; stars below are from cache-busted shields.io reads unless noted. Ranks 14 through 25 on the board were all brand-new repos offering "standalone" or "pre-configured" installers of paid software such as Adobe, Autodesk and Cubase, a classic malware lure, and are excluded.

  • aws/aws-emulator (HN, not on Trendshift): a local emulator of the AWS APIs for testing without an account, written by agents. Released today with a stream of v1.0.x fix releases (v1.0.322 on November 11). CC0-1.0 with an Apache-2.0 vendor/ folder, about 700 stars, telemetry on by default, no PRs accepted, and source builds want 100 GB of free disk.
  • meta/astryx (#5): Meta's open-source React design system built on StyleX and written to be queried by coding agents. MIT, about 25k stars, release feed not verified this run. Caveat: the README calls it beta, needs React 19 or later, and two packages ship only under a canary tag.
  • janestreet/agent_skills (#2): Jane Street's skills repo, currently one skill (socratic-tutor) that makes a model teach by asking questions. MIT, default branch master, no releases. Caveat: star readings disagreed between shields.io and the GitHub page this run, so treat the count as unknown; it is a one-skill repo riding a famous name.
  • klad-app/klad (Show HN): the agent-drivable paint app above. Under 100 stars (shields.io and the GitHub page disagreed), no releases. Caveat: PolyForm Internal Use license with a 500,000 euro revenue cap; only the file format, CLI docs, skills and client crate are MIT.
  • FrigadeHQ/yogurt (Show HN): a free, open-source meeting-notes app that records calls without adding a bot to the meeting, pitched against Granola. MIT, a couple hundred stars, release feed not verified. Runs on Apple Silicon macOS 14.4+, Windows 10 and Ubuntu 22.04.
  • HybridRobotics/supermimic (#7): a framework that trains humanoid robots to track whole-body motion with a single differentiable world model, built on mjlab. Apache-2.0, about 900 stars, no releases. Caveat: the README gives no benchmark numbers and needs a CUDA GPU plus access to a sibling research repo.
  • NeuraCryption/SPECTRA (HN, 21 points): a 4-bit compression and serving runtime for 1.5B to 4B models on 4 GB NVIDIA cards. Fourteen stars. Caveat: the README calls it open-source evaluation code but the LICENSE is a non-commercial research agreement, and every benchmark is self-run on one laptop with no scripts.

What actually matters from today's signal

Track the test harness, not the model. AWS's post makes the argument out loud: the engineer's job, it says, is to build "the specs, tests, and merge gates that decide when that work is correct." Everything with real builder value today sits in that sentence. A local AWS that agents can hammer for free, a sub-cent judge that can grade every agent step against a checklist, encrypted session logs you can replay, websites wrapped as callable tools so an agent stops guessing at buttons. If you are deciding where to spend effort this quarter, put it in four places: a local test environment your agent can run on its own, cheap automated checks on every agent action, a record of what each session did, and a short list of allowed tools.

The counter-signal is that the checking layer is being built by the same parties who sell the thing being checked. AWS's agents wrote the emulator and AWS's tests grade it, Google benchmarks its own judge, Pomerado runs its own race. Meanwhile the human side is pushing back: osssaysno.com now lists 22 open-source projects, including immich, Zig and Servo, that ban AI-generated contributions outright, and a single Claude Code release had to fix four separate security holes, three of them in its own permission checks. An agent that passes a vendor's emulator and a vendor's judge has proved it can satisfy the vendor. Keep one check in your pipeline that nobody selling you agents wrote.

Source access notes: Vendor pages, Hacker News (Algolia stories since about 22:00 UTC November 10 plus the front page), Hugging Face papers, Trendshift and repo files were read between about 15:10 and 15:45 ET on November 11. openai.com/news returned a fetch error and was skipped; blog.langchain.com redirected off-host and was not followed; blog.google's AI listing returned navigation only, so the Gemini Decisions post was read directly. The hn.algolia.com and trendshift.io endpoints were blocked from both shells and were read through the fetch tool instead. Release feeds for five repos were rate-limited, so their release status is marked unverified; the aws/aws-emulator release tag comes from its releases page. GitHub tree listings for the emulator's issues/ folders were refused by robots rules. Hacker News point totals are omitted because the list endpoint and the item endpoint disagreed. An adversarial fact-check pass (Sonnet 5.5) caught eight items, all fixed: the fidelity-score note pinned to the wrong release, an unsourced OpenAI comparison, the 9297 port tied to packaged installs instead of source builds, two HN point totals, the Maya timeline and model attribution, Klad's star count, and an overcount of permission-check bypasses. Codex's article fact checks later surfaced two corrections folded back here: AWS's News Blog calls AWS Emulator generally available and says it does not evaluate IAM or resource-based policies, and the Gemini pricing page does list Gemini Decision 1, so the earlier line saying it had no row was removed. Items covered in the last two briefings (VulnCheck's Glasswing tracker, Gemini 4 Carbon, Modern Web Guidance, Claude Whiteboards, Apple's reported "Harbor," Notion and Fireworks, AlphaProof Code) were skipped.