Agentic AI News Nov. 21, 2026

The day's AI agents news in one 5-minute read: releases, tools, research, funding and enterprise moves.

curated from 30 sources Saturday · Nov. 21, 2026 Back to today

Agents & Interfaces

Gemini runs Mac apps. Google's Gemini Mac app now has Computer Use, letting it complete multistep tasks across local files, apps, and websites in an isolated background window. Users choose which apps it can access and can approve key actions, with the feature rolling out in English in the US.

Apple's agent plans. Apple is hiring to build its own device-controlling AI agent, expected as early as late next year, while macOS 27.3 makes third-party agents like ChatGPT harder to enable. Bloomberg reports the new permission flow requires 14 steps instead of four, which some competitors see as a barrier.

Hermes Agent for Teams. Nous Research's open-source Hermes Agent will run inside Microsoft Teams, with a 'Collective Wisdom' feature that turns successful workflows into shareable internal skills. Agents run in isolated Azure Container Apps Sandboxes, and company data is kept out of AI provider training.

Google's AI search redesign. Google is rolling out a redesigned desktop search homepage that drops the dedicated search button and adds quick access to Gemini image generation, file analysis, and brainstorming. A new collapsible sidebar in AI Mode lets users reopen recent AI searches and start temporary threads.

Genesis AI robot learning. Genesis AI raised a $500 million Series A at a $3.5 billion valuation to build integrated robot learning, pairing a human-scale robotic hand with a data-collection glove, the GENE-26.5 model, Genesis World simulator, and Eno robot. The goal is to make unfamiliar tasks cheaper to teach and easier to trust.

Models & Tools

DaisyRose-27B coding model. RadixArk released DaisyRose-27B, an open-weight agentic coding model post-trained from Qwen3.8-27B using only reinforcement learning, with no SFT or distillation. It doubles Terminal Bench 4.1 pass@1 from 5.6 to 12.0 and ships under Apache 2.0.

Inkling 2 Mini. Thinking Machines Lab released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B total and 3B active parameters, a 1M-token context, and a looped transformer design. It targets local setups and supports images plus text, with NVFP4, MXFP8, and MLX support.

Nemotron-Labs-Sigma diffusion LLM. Nvidia released Nemotron-Labs-Sigma, a family of continuous diffusion language models at 3B and 8B that denoise text in a 16-D embedding space rather than discrete tokens. The 8B model averages 71.16 on Nvidia's six-task suite, and the 8-step distilled version retains a 70.34 average.

AWS emulator built by agents. AWS released a free, CC0 local emulator covering more than 200 AWS services, built entirely by autonomous AI agents. It runs as a single Rust binary on port 9297, so existing SDKs, CLIs, and coding agents can use it via environment variables.

Agent infrastructure to CNCF. Google donated Agent Substrate, gVisor, and OpenRL to the CNCF to secure AI agents at scale. The projects address isolated execution of untrusted AI-generated code, agent lifecycle management, and large-scale RL orchestration.

Harness adapters. Hugging Face introduced harness adapters, LoRA-style adapters trained with RL to optimize a model for a specific agent harness and task distribution. An example on Qwen3.8-27B lifts Terminal-Bench 2.1 pass@1 from 39% to 55% while cutting average token usage from 2.9K to 1.5K.

Security & Privacy

OpenAI reasoning safeguards. OpenAI says it sharply reduced extraction of protected reasoning from GPT-6.2 Astra: attacks on prior requests dropped from 54.0% to 0.2% and on current requests from 22.9% to 0.9%. The company is also sharing how it detects and disrupts adversarial distillation.

Signal blocks AI agents. Signal is restricting screenshots on iOS and Android to stop OS-level AI agents from reading private chats, calling the agents 'hostile to privacy.' It argues Apple, Google, and Microsoft lack easy ways for apps to block AI access without drastic compromises.

AI supply chain leaks. A researcher planted 2 million canary credentials in AI service requests and found sensitive data leaking through AI routers and gateways like OpenRouter and Vercel. The experiment exposed prompt theft, privacy violations, and exploited leaked AWS keys.

Research & Benchmarks

Riemann hypothesis claims. Anthropic mathematician Levent Alpöge says he and Claude have proved the Riemann hypothesis, with a Lean-compilable solution, while a separate all-human proof by Alessandro Masullo targets the Generalized Riemann Hypothesis. Both claims await independent verification.

METR time horizon update. METR's updated time horizon report finds some AI capabilities improving exponentially, with task length completed doubling every four months in earlier data, but models still lag human judgment tasks compared with coding. The new plot adds longer and more open-ended tasks.

MedChemBench for drug discovery. OpenAI introduced MedChemBench, a benchmark of 99 medicinal chemistry reasoning problems to test whether AI can connect molecular structures to experimental evidence and choose which compounds to try next.

Hebbia's financial research gains. Hebbia tested GPT-6 Astra in its Matrix financial research tool and found citation relevance of 99.5% and a 13.6-point lead over the best alternative on required source types. The model also met 97.9% of presentation requirements.

Formal verification versus slop. As AI-generated code raises the cost of low-quality output, startups including Harmonic, Theorem, Axiom Math, and &Truth are offering formal verification systems. They use mathematical proofs and SMT solvers like cvc5 to verify software correctness.

Business & Policy

Anthropic's record IPO. Anthropic closed the largest IPO in history, raising about $93.5 billion in gross proceeds and trading on Nasdaq under 'ANT' at a valuation near $1.5 trillion. Its Class F shares retain 50.1% voting power for the founders.

Nvidia's Inception acquisition. Nvidia is nearing a deal to acquire diffusion language model startup Inception for at least $500 million to strengthen open-source Nemotron models and physical robotics models. The startup's co-founders are expected to join Nvidia full-time.

Meta rethinks net-zero. Meta is re-evaluating its 2030 net-zero emissions goal, citing AI data center demand that grew faster than projected. The company reported a record percentage increase in greenhouse gas emissions, exceeding other big tech competitors.

Trump's data center tax. President Trump threatened a 'tremendous tax' on US companies that build AI data centers outside the country, arguing that China will take them if allowed abroad. He provided no details on the proposal.

AI jobs in China. A Bloomberg analysis of nearly 10 million Chinese job postings found AI-exposed roles fell about 10 percentage points more than less affected roles between mid-2021 and mid-2025. Entry-level and creative jobs are feeling the shift most.

Space data center financing. Lenders in SpaceX's $40 billion chip financing deal are worried about repossessing chips installed in orbit, since chips in data centers usually serve as collateral. The Financial Times reports lenders asked that none of the chips be used for orbital data center plans.

That's everything for this day – about a 5-minute read.