Vercel Connect GA and Run SDK, Wan 3.0 on AI Gateway, and agent tools for production
TL;DR
Vercel expanded agent connectors and security tools while new video models and debugging platforms rolled out, giving builders more ways to ship reliable agents without managing credentials or long timeouts.
What shipped
On 25 August 2026 hardware and platform updates arrived alongside agent-focused releases from Vercel and LangChain. Hardware like the Pixel 11 sits next to managed connectors and evaluation SDKs that reduce credential sprawl. Trending models and research papers added options for video generation and retrieval.
Vendor launches
Vercel supplied the largest share of releases with managed connectors, a secure execution SDK, and general availability of short-lived token auth. Google added Pixel 11 phones and intelligent dictation in Gemini for macOS. NVIDIA and other vendors contributed smaller updates around gaming and training programs.
- •Vercel Connect Linq support Vercel added a managed Linq connector so agents can send messages over iMessage, RCS, and SMS without storing credentials.
- •Run SDK release Vercel shipped the Run SDK to execute untrusted JavaScript and TypeScript safely with built-in pauses for human approval or auth steps.
- •Next.js security protection Vercel shielded hosted Next.js apps from the August 2026 vulnerabilities so customers needed no extra fixes.
- •Vercel Connect GA Vercel made Connect generally available on all plans, replacing long-lived secrets with runtime-scoped tokens for any registered service.
- •AI Gateway async video AI Gateway now supports asynchronous video generation so apps can receive webhooks instead of holding open long HTTP requests.
- •Wan 3.0 on AI Gateway Alibaba Wan 3.0 became available as alibaba/wan-v3.0-video for text-to-video and reference-based clips up to 30 seconds.
- •MiniMax free tier MiniMax M3 and M2.7 models are free on AI Gateway via GMI Cloud until September 6 for testing coding agents.
- •Credential sprawl end Vercel Connect grew past 100 connectors during beta, letting agents reach Slack, Snowflake, and APIs with expiring tokens.
- •Gemini dictation on macOS Google enabled intelligent dictation in the Gemini app so users can speak into any desktop window.
- •Delaware AI training Google partnered with Delaware to offer free Career Certificates and AI training to state residents.
- •Notion Chat SDK adapter Vercel released a Notion adapter so agents can reply in page comments and threads without new code.
Product Hunt picks
Two small tools appeared on Product Hunt for ad transparency and session search. They sit alongside larger platform releases but target specific daily workflows.
- •ChatGPT Ad Library The library shows every ad currently running inside ChatGPT for marketers tracking competitor placements.
- •session-indexer The tool adds semantic search across a user's own Claude Code session history for quick recall of past prompts.
Hugging Face trending
Four models and papers trended on Hugging Face. Two new checkpoints focus on multimodal and uncensored text, while research papers address 3D rendering fixes and road safety audits.
- •SenseNova-U1.5-8B-MoT The any-to-any model from sensenova is available on the Hub for download, fine-tuning, and inference with transformers.
- •Qwen3.8-27B-Uncensored The image-text-to-text model from orcarouter is trending and ready for local or Hub-based inference.
- •FixAnything paper The work uses video generative priors to refine 3D-consistent rendering from sparse input views in Gaussian Splatting or NeRF scenes.
- •EG-ARSA paper The expert-grounded model supports visual road safety auditing in low-resource settings where crash records are incomplete.
Other
LangChain posted the largest set of updates covering LangSmith improvements, agent skills, retrieval patterns, and SQL workflows. Additional posts from Lambda and EleutherAI addressed agent learning and deception detection.
- •LangSmith Engine update LangSmith Engine now detects agent issues over twice as well and proposes fixes that integrate with Slack and Linear.
- •Deep Agents skills LangChain showed how to load and run skills dynamically inside the Deep Agents CLI for token-efficient agents.
- •Aletheia Quest retrospective EleutherAI shared lessons from building black-box and white-box detectors for AI deception.
- •Web research automation LangChain demonstrated parallel searches and page scraping with retrievers to synthesize results locally or in the cloud.
- •Plan-and-Execute agents LangChain outlined a framework that separates planning from execution to reduce errors on complex tasks.
- •LLMs and SQL guide LangChain explained techniques to query databases in natural language while cutting hallucinations.
- •Context engineering rise LangChain described building dynamic systems that supply LLMs the right tools and format for reliable output.
- •Deep Agents debugging LangChain covered tracing with LangSmith and prompt optimization using Polly for complex agent traces.
- •Structured Tools guide LangChain explained how to accept multiple inputs and create complex schemas for advanced agent capabilities.
- •Multi-Vector Retriever LangChain provided cookbooks for retrieving across tables, text, and images in semi-structured RAG setups.
- •AgentFlow workflow learning Lambda described systems where the agent's workflow itself learns instead of remaining hand-wired.
- •CSV Q&A benchmarking LangChain released benchmarks and code for reliable question answering over CSV data with agents and retrieval.
- •Timescale Vector integration LangChain highlighted 243 percent faster similarity search and time-based RAG using PostgreSQL.
- •LangServe Playground LangChain added a configurable playground UI for deploying apps and streaming responses in real time.
- •Flexible retrieval methods LangChain reviewed semantic, hybrid, and custom retrievers for building personalized ChatGPT-style apps.
- •Cube semantic layer LangChain showed how the semantic layer prevents hallucinations when querying data conversationally.
Replicate new models
Replicate added two models from ByteDance and IBM for image generation and reasoning tasks that builders can call directly through the existing API.
- •seedream-5-pro on Replicate ByteDance released seedream-5-pro for text-to-image and image editing at 1K or 2K resolution from text or up to ten references.
- •granite-4.2-8b on Replicate IBM Granite 4.2 8B landed as a mid-size reasoning model suited for tasks that need strong chain-of-thought performance.
Industry news
OpenAI blocked a Russian influence operation that used ChatGPT. MIT research covered AI in caregiving startups, cancer vaccine response, underwater vision systems, and how chatbots affect news accuracy.
- •Russian ChatGPT campaign OpenAI banned accounts that used ChatGPT to generate pro-Kremlin social posts through VPNs from Russia.
- •AgeLab AI startup An MIT AgeLab study on caregiving tech prompted a serial entrepreneur to launch a simpler device startup.
- •Brain on AI study MIT Media Lab research found that repeated use of AI chatbots for news lowered participants' headline accuracy over four weeks.
What this means for you
For Vibe Builders: You can now connect agents to iMessage and Notion without writing auth code using Vercel Connect and the new Chat SDK adapter. Free MiniMax models and async video generation on AI Gateway let you test longer workflows without hitting timeouts or paying upfront. LangSmith's improved issue detection and the Run SDK give you safer ways to run agent code while keeping human approval steps.
For Non-techies: Google's intelligent dictation in Gemini for macOS and the Pixel 11 phones make everyday desktop and mobile tasks faster without extra setup. Free tiers on Speed Insights and MiniMax models on AI Gateway reduce costs when you want performance data or quick video tests. Delaware residents can access free AI training certificates through the new Google partnership.
For Developers: Vercel Run SDK and Connect GA replace long-lived tokens with scoped runtime credentials, so evaluate them against your current agent auth stack. LangSmith Engine's 2x better issue detection and the new multi-vector retriever patterns give concrete benchmarks for production RAG and agent reliability. Watch the AgentFlow and context engineering posts for signals on moving from hand-wired workflows to learned ones.
What to watch next
Track adoption numbers for Vercel Connect connectors and the Run SDK over the next seven days. Watch for production benchmarks comparing Wan 3.0 against earlier video models on AI Gateway. Monitor LangSmith self-hosted releases for any reported reliability gains in agent issue detection.
Harsh’s take
The day showed heavy concentration around Vercel agent infrastructure while model releases stayed incremental. Many updates still require developers to stitch together connectors, SDKs, and tracing tools rather than offering one integrated path. The practical move this week is to pick one connector or the Run SDK, wire it into an existing agent, and measure credential and timeout reductions against your current setup before adding more pieces.
by Harsh Desai
Sources
Vendor launches
- •You can officially buy the Pixel 11 phones and Pixel Watch 5.
- •Vercel Connect now supports Linq
- •Introducing Run SDK: secure eval for your agents
- •Vercel applications are protected from Next.js August 2026 security vulnerabilities
- •Vercel Connect is now generally available
- •AI Gateway now supports asynchronous video generation
- •Wan 3.0 now available on AI Gateway
- •MiniMax M3 and M2.7 are free on AI Gateway
- •The end of credential sprawl for agents
- •Leading Publishers Bring Blockbuster PC Games and Technology to NVIDIA RTX Spark
- •Here’s how to use intelligent dictation in Gemini for macOS.
- •Speed Insights adds a free tier
- •We’re partnering with the State of Delaware to provide free AI and career training.
- •Bring your agent to Notion with Chat SDK
Product Hunt picks
Hugging Face trending
- •SenseNova-U1.5-8B-MoT by sensenova trends on HuggingFace
- •Qwen3.8-27B-Uncensored by orcarouter trends on HuggingFace
- •FixAnything: 3D-Consistent Rendering Refinement via Video Generative Priors
- •EG-ARSA: An Expert-Grounded Open Model for Visual Road Safety Auditing in Low-Resource Settings
Other
- •LangSmith Engine Improves Agent Issue Detection by 2x
- •Using skills with Deep Agents
- •What We Learned Trying to Catch AI Liars: An Aletheia's Quest Retrospective
- •Automating Web Research
- •Plan-and-Execute Agents
- •LLMs and SQL
- •The rise of "context engineering"
- •Debugging Deep Agents with LangSmith
- •Structured Tools
- •Multi-Vector Retriever for RAG on tables, text, and images
- •AgentFlow: when the agent's workflow learns
- •Benchmarking Question/Answering Over CSV Data
- •Timescale Vector x LangChain: Making PostgreSQL A Better Vector Database for AI Applications
- •LangServe Playground and Configurability
- •Retrieval
- •Cube x LangChain: Building AI experiences with LLMs and the semantic layer
Replicate new models
Industry news
More AI news
- Daily RoundupAudio8 TTS and Wan 3 video model launch, plus agent efficiency benchmarks (sandbox tools to test today)
Hugging Face and Replicate added new text-to-speech, video, and image models while NVIDIA and Vercel released efficiency and sandbox updates that affect how agents run in production.
- Weekly DigestHermes Agent v0.20 voice and A2A launch, OpenClaw GPT-5.6 support, and agent stability fixes (two flagships + practical hook)
Hermes Agent rolled out v0.20 series with conversational voice, A2A protocol, and desktop UI updates while OpenClaw added model support, plugin safeguards, and mobile caching across the week.
- Daily RoundupReplicate voice clone and turbo image model, Ornith-1.5 trends, Ox Alpha speculation
Replicate added a zero-shot music voice cloning tool and a fast 6B-parameter image LoRA while Ornith-1.5-9B gained traction on Hugging Face and a stealth model called Ox Alpha stirred online talk.