Muse agent tops charts, Vercel adds sandbox memory tools, and skills reach one million installs

What shipped

On 25 September, Meta accelerated promotion of its Muse agent across its apps as the system gained users rapidly. Vercel shipped memory observability for sandboxes plus GitHub Actions support for its container registry. Model hosts released new text-to-image, sound, and instruction models while industry reports covered large cloud contracts and security rulings.

Vendor launches

Vercel supplied all four updates in this section. The company added memory usage cards and CLI metrics to its sandbox product, introduced an OIDC-based login action for the container registry, published adoption numbers from the skills.sh registry, and detailed how Klaviyo enabled 512 employees to ship 356 private apps in two weeks.

• •Vercel Sandbox memory metrics Vercel added average, P75, and P95 memory readings to the sandbox dashboard and the vercel metrics command, letting teams compare usage against existing CPU and transfer numbers before scaling workloads.
• •Vercel Container Registry GitHub action The new vercel/vcr-action/login action swaps GitHub OIDC tokens for short-lived Vercel tokens so teams can push images from workflows without storing long-lived registry keys.
• •skills.sh registry growth The registry reached one million agent skills and 280 million installs in seven months, giving teams ready-made instruction files that turn generic agents into role-specific workers without extra code.
• •Klaviyo internal apps on Vercel Klaviyo opened a Vercel-based platform to 512 employees and reached 356 live, SSO-gated apps in two weeks, cutting the time from idea to private app to three minutes per builder.

Hugging Face trending

Hugging Face listed five trending models. Four are ready-to-run checkpoints for image-text, text-to-image, text generation, and classification tasks, while one research paper describes compiling agent skills into state machines for more reliable control flow.

• •TeleOCR image-text model StarDoc-AI released TeleOCR, an image-to-text model on the Hub that teams can download or fine-tune for document transcription jobs without hosting their own inference stack.
• •Qwen-Image-2.1-GGUF text-to-image model unsloth published a quantized Qwen image model that runs locally or on the Hub, giving vibe builders a drop-in replacement for slower text-to-image calls.
• •MiMo-V2.6-Flash-RL text model XiaomiMiMo added a reinforcement-learned text-generation checkpoint that developers can test directly on the Hub for faster instruction following than prior versions.
• •Jev-Omni classification model akhilaaa3 released Jev-Omni, a text-classification model on the Hub that agents can call to label incoming messages before routing them to the next skill.
• •HEXIS skill compiler paper The HEXIS paper shows how to turn plain-language agent skills into finite state machines so agents follow prescribed steps without skipping or repeating actions.

Fal model gallery

Elevenlabs Sound Effects V2 on Fal: Elevenlabs Sound Effects V2 is now callable on Fal for text-to-audio tasks, letting teams generate sound clips for video or agent responses without managing separate audio infrastructure.

Product Hunt picks

Nine tools appeared on Product Hunt. They cover token forecasting, civil engineering design, multi-user test agents, agent build galleries, React UI kits, video editing by chat, AI filmmaking, performance marketing automation, and carbon tracking for tokens.

• •Token Forecaster The tool predicts token counts for an LLM reply before the call runs, helping teams avoid surprise bills on long outputs.
• •Jango Jango spins up AI agents that mimic real users to test multi-user apps, replacing manual QA scripts with repeatable agent sessions.
• •JevForAgents The gallery collects real Jev agent builds and demos so builders can copy patterns instead of starting from blank prompts.
• •Once UI 2.0 The updated React kit produces consistent interfaces that both human developers and AI agents can edit, reducing styling drift across projects.
• •SocialGPT The app lets users edit timeline videos by typing instructions, removing the need to learn separate video editing software.
• •Hyperdream The tool positions itself as the Cursor equivalent for AI filmmaking, letting creators iterate on scenes through prompts rather than timeline edits.
• •HireOtto The platform runs an entire performance marketing stack through AI agents, collapsing campaign setup, bidding, and reporting into one chat interface.
• •CrbonFree The service produces audit-grade carbon figures per AI token so teams can track emissions alongside usage costs.

Other

Six releases appeared outside the main categories. They include an AWS throughput benchmark, a GitHub Copilot canvas tutorial, two OpenRouter model additions, a LangGraph and Jev production guide, and a duplicate Perceptron listing.

• •AWS MoE reinforcement learning scaling AWS reported 40 percent higher throughput when running mixture-of-experts reinforcement learning on EKS with EFA and DeepEP, giving teams a concrete cluster recipe for large training jobs.
• •GitHub Copilot canvases for beginners The post shows how to describe a needed interface in plain English so the agent builds a live, editable surface inside the editor.
• •Perceptron Mk1.5 on OpenRouter The embodied reasoning model accepts text, image, video, and audio and returns structured outputs such as points and polygons for physical agent tasks.
• •Jev and LangGraph production agents The LangChain post demonstrates how LangGraph orchestrates Jev decision models to cut cost and latency in deployed agent systems.
• •Perceptron Mk1.5 pricing details OpenRouter lists the model at 37k context with $0.15 per million input tokens and $1.50 per million output tokens for quick testing.
• •Jev Router on OpenRouter The router selects the best model and reasoning effort per request and runs on the Jev System One model with a 1,000k context window.

Industry news

Twelve stories covered security incidents, model milestones, corporate moves, and policy rulings. Meta's Muse received the most attention for its rapid growth and persistent Linux VM architecture, while Anthropic signed an $11.6 billion cloud deal and faced a upheld Pentagon supply-chain restriction.

• •Supabase data exposure cases TechCrunch reported that several AI-generated apps left user data publicly accessible, showing the security risks when vibe-coded projects skip proper configuration.
• •Astra and Opus codebreaking milestone Frontier models completed remaining World War II codebreaking tasks that Turing's team left unfinished, providing a new historical benchmark.
• •Meta Muse app growth Muse climbed app-store charts while Meta expanded promotion, giving users their own persistent Linux VM in the cloud for agentic work.
• •Microsoft Copilot Autopilot agent Microsoft split Copilot into Home, Code, and a new always-on Autopilot agent built on OpenClaw with usage-based billing for the agent tier.
• •John Gruber on Muse Gruber noted that Muse is the first consumer agentic system packaged as an easy mascot app, yet consumers may not grasp the persistent VM implications.
• •Meta Muse overtakes model news TechCrunch observed that Muse grabbed attention from OpenAI and Anthropic releases because of its consumer packaging and early usage numbers.
• •Meta Muse Tamagotchi comparison Coverage framed Muse as a Tamagotchi-style persistent agent whose bet appears to be working despite questions about user understanding.

Replicate new models

qwen3-235b-a22b-instruct-2,507 on Replicate: The updated Qwen3 model improves instruction following and is available for direct calls via Replicate's HTTP API or existing tokens.

What this means for you

For Vibe Builders: You can now add memory checks to Vercel sandboxes before shipping internal tools, pull trending image and audio models from Hugging Face and Fal for quick prototypes, and copy Jev or HEXIS patterns to make agents follow steps reliably. Muse shows that persistent agents packaged for non-coders are gaining users fast, so test one on a small workflow this week. Watch the skills.sh registry for ready-made instruction files that cut prompt engineering time.

For Non-techies: For your business, tools like Klaviyo's three-minute app builder and HireOtto's marketing agents move AI from chat to finished tasks you can assign to staff without coding. Muse offers a simple mascot interface that runs its own VM, but check data settings first after the Supabase exposure reports. Token Forecaster and CrbonFree give quick ways to control spend and emissions on daily AI use.

For Developers: Production teams should benchmark the new Qwen instruction model and Perceptron Mk1.5 against current stacks, then test the Vercel OIDC action and AWS EKS scaling recipe for throughput gains. LangGraph plus Jev examples show concrete ways to orchestrate reliable agents, while the HEXIS compiler paper points to state-machine approaches that reduce skipped steps. Track the Anthropic cloud contract and Muse VM details for architecture signals before committing to new runtimes.

What to watch next

Track whether Muse reaches smart glasses integration and how OpenAI and Anthropic respond with their next releases. Watch for more teams publishing skills.sh entries and any follow-up on the Supabase exposure cases. Note new Replicate and OpenRouter model additions for cost or context improvements.

Sources

Vendor launches

• •Vercel Sandbox now supports memory observability
• •Push images to Vercel Container Registry from GitHub Actions
• •State of agent skills
• •How Klaviyo shipped 356 internal apps in two weeks on Vercel

Hugging Face trending

• •TeleOCR by StarDoc-AI trends on HuggingFace
• •Qwen-Image-2.1-GGUF by unsloth trends on HuggingFace
• •MiMo-V2.6-Flash-RL by XiaomiMiMo trends on HuggingFace
• •Jev-Omni by akhilaaa3 trends on HuggingFace
• •HEXIS: Compiling Skills into Extended Finite State Machines

Fal model gallery

• •Elevenlabs Sound Effects V2 drops on Fal

Product Hunt picks

• •Token Forecaster
• •InfraGrid3D
• •Jango
• •JevForAgents
• •Once UI 2.0
• •SocialGPT
• •Hyperdream
• •HireOtto
• •CrbonFree launches on Product Hunt

Other

• •Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
• •GitHub Copilot app for Beginners: How to build custom workflows with canvases
• •Perceptron: Perceptron Mk1.5 now available on OpenRouter (37k context, $0.15/M in, $1.50/M out)
• •Building Production Agents with Jev and LangGraph
• •TypeSafe: Jev Router now available on OpenRouter (1,000k context, $-1,000,000.00/M in, $-1,000,000.00/M out)
• •TypeSafe: Jev Router now available on OpenRouter (1,000k context, $-1,000,000.00/M in, $-1,000,000.00/M out)

Industry news

• •Some Supabase customers are publicly exposing reams of people’s data to the web
• •Astra and Opus just passed Turing’s other test
• •Meta is putting its muscle behind Muse as the AI app takes off
• •Appeals Court Lets the Pentagon Designate Anthropic a Supply-Chain Risk
• •Another Google Deepmind researcher quits, says building superintelligent AI soon is "inherently irresponsible"
• •Microsoft gives Copilot another makeover, adding an Autopilot agent and usage-based billing
• •Quoting John Gruber
• •Anthropic to pay Akamai $11.6 billion over seven years in cloud deal
• •Ahead of US IPO, British AI neocloud Nscale secures $3.36B in convertible financing
• •Meta’s Muse just stole the AI spotlight from OpenAI and Anthropic
• •Meta’s AI Tamagotchi bet is…working?
• •Pentagon was right to slap Anthropic with a security supply chain risk label, federal court says

Replicate new models

• •qwen3-235b-a22b-instruct-2,507 dropped on Replicate today