Claude opens Duet preview; OpenAI models are coming to Google Cloud

#1 𝕏

Claude opened a research preview of Claude Duet, a new app where Claude acts as a proactive agent. It surfaces items needing attention, completes requested tasks, continues working when users step away, and checks before acting on their behalf.

Also covered by: @Mike Krieger, @Dan Shipper, @Mike Krieger, @How I AI Podcast, @claire vo 🖤, @claire vo 🖤

#2 📝 OpenAI News

OpenAI models are coming to Google Cloud - OpenAI News: On October 21, 2026, OpenAI announced that its models are coming to Google Cloud through a partnership involving OpenAI’s frontier models and Codex; the article has a 1:46 listening time.

Also covered by: @OpenAI

#3 𝕏

xAI shared that Grok Bot is its always-on AI teammate, with each bot running on a persistent cloud computer equipped with a browser, files, and terminal. The bots can learn workflows shown once, execute multi-step tasks 24/7 even after a user closes their laptop, and coordinate in parallel.

Also covered by: @Peter Yang, @Grok

#4 𝕏

Claude Security connects to repositories once, performs a deep first scan, and re-scans as changes land to keep findings close to the current code state. It is in public beta on Claude Enterprise.

#5 📝 OpenAI News

Patch the Planet: global response coordination as agents uncover critical vulnerabilities - OpenAI News recapped how, over the preceding two months, OpenAI led a coalition of maintainers, researchers, and security providers to address critical vulnerabilities in open-source PHP.

Also covered by: @openai/php2shell-checker, @OpenAI

#6 𝕏

Google Research said it completed the 32-bit open-source Coral NPU architecture, delivering up to 1 TOPS under 10 mW, and released its v1.0 compiler. It also launched new Coralboard development kits with Synaptics.

#7 𝕏

Mistral AI announced a collaboration with NVIDIA to bring AI to industrial engineering, covering faster simulation, open benchmarks and supporting datasets for physics AI, and evaluations for agentic AI in CAE and EDA.

#8 🐙 ollama/ollama

ollama/ollama - **ollama/ollama** is an MIT-licensed app with a CLI, REST API, Python and JavaScript libraries, Docker image, documentation, and integrations for running and managing open models locally or in the cloud. On October 21, 2026, Ollama released v0.40.8, adding Recursive’s Flint coding model with multi-turn tool use and a 262K-token context window on macOS. Worth trying if you need an open-model stack for coding agents. Release notes

#9 𝕏

Harrison Chase shared LangChain’s SmithTune approach: use successful logged traces to fine-tune a small model and evaluate it on the target task. He said a 9B model can beat frontier models when trained on exactly that task, though no benchmark details were provided.

#10 🐙 akitaonrails/ai-memory

akitaonrails/ai-memory - akitaonrails/ai-memory is an MIT-licensed long-term memory tool for coding agents; owner akitaonrails released v3.2.0 on October 22, 2026. Its binary, hooks, server, and git-backed Markdown wiki share context and handoffs across more than 20 agent harnesses with zero LLM calls by default and a stated ceiling of approximately 700 writes per second; the repo has 12,605 stars, up 286 daily and 2,833 weekly. Consider it if you need context across agents or machines. Release notes

#11 🐙 heygen-com/hyperframes

heygen-com/hyperframes - heygen-com/hyperframes is an Apache-2.0 framework for turning HTML, CSS, media and seekable animations into deterministic MP4 videos, with a CLI, 7 agent skills and about 450 editable blocks and components. heygen-com released v0.8.185 on 2026-10-21, adding 4K transparent MOV and WebM cloud rendering plus opt-in motion blur; the repo lists 69576 stars and 5740 weekly trending gains. Worth trying if you need agents to build HTML-based video and render it locally or in the cloud. Release notes

#12 🐙 vllm-project/vllm

vllm-project/vllm - vllm-project/vllm shares vLLM, an Apache-2.0 library for LLM inference and serving, with OpenAI-compatible and Anthropic Messages APIs, support for 200+ Hugging Face model architectures, plus source code, documentation, examples, benchmarks, and recipes. Worth trying if you need flexible model serving across supported GPUs, CPUs, and hardware plugins.

#13 🐙 langfuse/langfuse

langfuse/langfuse - **langfuse/langfuse** is an open-source LLM engineering platform for tracing, prompt management, evaluations, datasets, and debugging, with skills and an AI gateway in the repository. Langfuse released v4.67.0 on October 21, 2026, adding simpler skills setup and API-key filtering alongside multiple fixes; the project has 36,174 GitHub stars. Consider it if you need a self-hostable platform for monitoring and improving AI applications. Release notes

#14 🐙 n8n-io/n8n

n8n-io/n8n - **n8n-io/n8n** is a fair-code platform for building AI agents and workflows with a visual canvas, custom code, self-hosting or cloud deployment, 1500+ integrations, and 9,000+ templates. n8n-io released n8n@2.44.6 on October 22, 2026, fixing a Postgres Node issue that could overwrite global pgtypes. Consider it if you need flexible AI workflow automation across models and tools. Release notes

#15 🐙 langchain-ai/langchain

langchain-ai/langchain - **langchain-ai/langchain** is an MIT-licensed framework for building agents and LLM-powered applications, with libraries of interoperable components and third-party integrations. On 2026-10-21T20:43:48Z, langchain-ai released langchain-core==1.6.13, adding support for reading `tool-call-delta` in stream consumers and bumping certifi in `/libs/core` from 2025.11.12 to 2026.10.19. Consider it if you need reusable building blocks for agent or LLM applications. Release notes

#16 🐙 langgenius/dify

langgenius/dify - langgenius/dify is an open-source platform for building visual AI workflows, RAG pipelines, and agents with model management and observability. The repository includes the app, API, CLI, agent runtime, Docker setup, documentation, and end-to-end tests. Worth trying if you need one workspace for developing and deploying LLM applications.

#17 🐙 Tencent/AI-Infra-Guard

Tencent/AI-Infra-Guard - Tencent/AI-Infra-Guard is an Apache-2.0 AI red-teaming platform from Tencent Zhuque Lab, with a web interface, CLI tools, and agent, skill, MCP, infrastructure, and jailbreak scanners. Tencent released v4.6.7 on October 22, 2026, with 4 additions and 6 fixes; the repository had 6,936 stars and 7 listed contributors. Worth trying for internal AI security checks, but it lacks authentication and should not be deployed publicly. Release notes

#18 🐙 comet-ml/opik

comet-ml/opik - comet-ml/opik is an Apache-2.0-licensed, self-hostable platform for LLM app and agent tracing, evaluation, and monitoring; it includes a backend, web app, SDKs, datasets, dashboards, guardrails, and optimization tools. comet-ml released Opik 2.2.115 on October 22, 2026, updating automatic OpenAPI specification and Fern code generation, defaulting OpenAI and OpenRouter online-evaluation models to GPT 5.5, and fixing the catalog so OpenAI reasoning models use at least 16 maximum output tokens. Consider it if you need self-hosted LLM observability. Release notes

#19 🐙 mlflow/mlflow

mlflow/mlflow - mlflow/mlflow is an Apache-2.0 platform with source for tracing, evaluation, monitoring, prompt optimization, model tracking, registry, and deployment. mlflow released MLflow 3.17.2 on October 21, 2026, adding dataset, scorer, and agent lineage to `mlflow.genai.evaluate` runs and fixing Spark and Optuna tracking issues. Worth trying if you need unified tooling for agents, LLMs, and machine-learning models. Release notes

#20 in

Guillermo Rauch announced an effort to extend the open-source, runtime-agnostic durability infrastructure powering eve and Workflows to agent harnesses such as Pi and OpenCode, with local, Vercel, or DIY usage. He said it can also make Claude Code and Codex durable and observable, though they are not currently designed for embedding, making them less efficient and portable and requiring an additional sandbox compared with Pi or Fx.