Skip to content
MyAI Guide
Newsletter
GLM 5.2 lands on Vercel and OpenRouter, Android 17 ships, and agent tools hit Product Hunt | Daily AI roundup cover
Daily RoundupIndustryVibe BuilderNon TechnicalDeveloper

GLM 5.2 lands on Vercel and OpenRouter, Android 17 ships, and agent tools hit Product Hunt

By Harsh Desai
Share

TL;DR

GLM 5.2 gains wide access for long tasks while Android 17 and dozens of agent tools push practical AI use forward on 16 June.

What shipped

On 16 June multiple vendors released updates that extend model reach and agent capabilities. Vercel added longer runtimes and GLM 5.2 support while Google pushed Android 17 and Pixel features. Product Hunt listings and platform posts showed a surge in tools for running agents at scale.

Vendor launches

Vercel led with three updates that directly affect long-running agent work. GLM 5.2 arrived on its AI Gateway with a 1M token window, sandboxes extended to 24 hours, and the Workflow SDK gained TanStack Start support. Other releases included HPE and NVIDIA hardware expansions plus Google device features tied to Gemini.

  • •Vercel Sandbox Sessions now run up to 24 hours on Pro and Enterprise plans, supporting large data jobs and long agent workflows that previously hit the five-hour limit.
  • •UK housing AI Google DeepMind and the UK government released a prototype that speeds planning decisions for new house building projects.
  • •GLM 5.2 on AI Gateway The model is live with a 1M token window and better long-task reliability; set zai/glm-5.2 in the Vercel AI SDK to test project-scale engineering work.
  • •HPE AI Factory NVIDIA and HPE expanded the factory offering at Discover Las Vegas with Vera CPU and Agent Toolkit support aimed at moving agents from test to production.
  • •June Pixel Drop Google added screen recording, Gemini Omni text-to-video, and improved multitasking to Pixel phones.
  • •Wear OS 7 The update brings glanceable widgets and later Gemini Intelligence to select smartwatches.
  • •NVIDIA Blackwell The platform topped MLPerf Training 6.0 benchmarks, giving teams faster iteration on large model training runs.
  • •Workflow SDK TanStack Start apps on Vercel can now run durable workflow and step functions written in plain TypeScript.

Replicate new models

Three new models appeared on Replicate for direct inference. One handles face swaps, another offers fast image generation, and a third converts scanned PDFs and tables into clean Markdown plus layout JSON.

  • •face-swapper The naimish-gami model runs face swaps on Replicate with options to enhance results from input images.
  • •vishnu The 55zn56bk8t-art model accepts prompts and seeds for quick image generation with a fast mode toggle.
  • •layoutgrok-qwen3-vl-cp500 This fine-tuned Qwen3-VL model turns complex PDFs and multi-column reports into Markdown and layout JSON on Replicate.

Hugging Face trending

GLM-5.2 from zai-org led trending models on the Hub. Two smaller models also gained attention for text and image-text tasks.

  • •GLM-5.2 zai-org's text model is now downloadable and fine-tunable on Hugging Face with documented evaluation results.
  • •VibeThinker-3B WeiboAI's 3B text model is trending and ready for local inference or further training via the Hub.
  • •MiniMax-M3-GGUF unsloth's image-text-to-text model is available in GGUF format for download and local use.

Product Hunt picks

Twenty-three distinct AI tools launched on Product Hunt, most focused on agents that hire, code, or run in sandboxes. Several target local or terminal workflows while others add support features like lead generation and customer chat.

  • •AGIRAILS Agents can now hire and pay each other with on-chain settlement for automated team tasks.
  • •Annota A local-first note app uses on-device AI for cross-platform capture and search.
  • •Fluxmail The tool sorts urgent email and drafts replies inside existing inboxes.
  • •YoAmigo Studio Builders deploy real apps using models they already subscribe to.
  • •Kepler An agentic environment scales multiple coding agents across projects.
  • •Avocado AI content tools update any Next.js site without leaving the editor.
  • •GlassKit UI React components prepare apps for Meta Ray-Ban smart glasses.
  • •Amaroad Slide decks are generated from the terminal using AI prompts.
  • •Voice Calls in Chatwoot Support inboxes now combine calls, chat, and email threads.
  • •HTML Deployer AI chat output converts directly to live HTML sites.
  • •Tychi AI Autonomous capital systems manage economic decisions for agents.
  • •Canopy Sandboxed Claude Code sessions run in parallel on macOS.
  • •Athena Desktop A local dashboard controls multiple AI coding agents.
  • •agentbrowse Coding agents gain web access through a command-line interface.
  • •whoburnedmore Public leaderboards track token spend across Claude and Codex sessions.
  • •Edgee Turbo Models Claude Code users gain access to Kimi K2.7 and MiniMax variants.
  • •ClientJam Designers receive AI-generated leads for new projects.
  • •Botme A customer support agent appears on any site in five minutes.
  • •Sklm Skills are scoped and synced across every agent in a workspace.
  • •ResumeWriting.com The service focuses solely on AI resume generation.
  • •Notum Law firms gain AI research and document intelligence tools.

Other

LangChain, Databricks, AWS, and OpenRouter posted updates that improve agent evaluation, data sharing, and model access. Several posts focused on feedback loops and governed workflows for production agents.

  • •P-EAGLE on SageMaker AWS added parallel speculative decoding to speed inference on its AI platform.
  • •OpenSharing on Databricks Delta Sharing evolved into a governed layer for agentic data exchange.
  • •Apps on Databricks Marketplace Teams can now publish and discover data intelligence apps inside the marketplace.
  • •LangSmith Benchmarks Public datasets let users compare RAG and agent performance across shared tests.
  • •LangSmith homepage The site now groups tools into Observability, Evaluation, and Prompt Engineering sections.
  • •Agent Engineering LangChain framed the new discipline that blends product, engineering, and data work for reliable agents.
  • •Fine-tuned models in LangSmith Users can test multiple open-source fine-tunes and pick the strongest performer.
  • •Human judgment loop LangChain showed how documented team knowledge improves agent accuracy over time.
  • •Loop Engineering Stacked agent loops with LangChain primitives raise reliability on real tasks.
  • •Fleet agents General chat and specialized agents run side by side for quick and recurring work.
  • •GLM 5.2 on OpenRouter The model offers 1,049k context at $1.40 per million input tokens.
  • •Governed vibe coding Databricks added controls so enterprise teams can run AI coding agents safely.
  • •Factory LangSmith case Iteration speed doubled after the team automated feedback loops in LangSmith.
  • •Monte Carlo agent LangGraph and LangSmith helped build an AI troubleshooting agent for data teams.
  • •SageMaker container caching AWS reduced model scaling time with new caching on its AI service.
  • •Bedrock Guardrails API Amazon added an invoke check that protects agentic apps from unsafe output.

Industry news

Google released Android 17 alongside Wear OS 7 and expanded Gemini features across devices. Separate reports covered consumer pushback on AI branding and pricing shifts at Microsoft Copilot.

  • •Android 17 Google launched the OS with new multitasking tools, parental controls, and deeper Gemini integration on Pixel hardware.
  • •AI brand survey Sixty percent of US consumers said seeing AI in marketing messages reduces trust in the brand.
  • •Qwen3.6-27B feedback Georgi Gerganov reported daily use of the local model for routine coding tasks on M2 Ultra and RTX hardware.
  • •Copilot Cowork billing Microsoft moved the product to usage-based pricing and is testing a fine-tuned DeepSeek V4 option.

What this means for you

For Vibe Builders: You can now run longer agent sessions with Vercel Sandbox at 24 hours and test GLM 5.2 directly in the AI SDK without managing infrastructure. Product Hunt listings give ready-made starters for hiring agents, local note tools, and terminal slide builders that drop into existing workflows. Watch the new Workflow SDK examples to move a prototype into a durable, restart-proof process this week.

For Non-techies: Android 17 and the Pixel Drop bring screen recording plus Gemini text-to-video straight to phones and watches you already own. Several Product Hunt tools let you add AI email sorting or website chat support in minutes without code changes. Fitbit Air purchases now qualify for pre-tax health accounts if you track daily routines.

For Developers: GLM 5.2 on Vercel AI Gateway and OpenRouter gives a 1M token window for project-level tasks that older 200K models could not handle. LangSmith public benchmarks and the new agent engineering posts supply concrete datasets to compare your current stack against fine-tuned open models. Evaluate P-EAGLE speculative decoding on SageMaker and container caching before the next production rollout.

What to watch next

Track GLM 5.2 adoption numbers on OpenRouter and Vercel this week. Watch for early user reports on 24-hour Vercel Sandboxes with persistent state. Note any new Product Hunt launches that integrate with the Workflow SDK or LangSmith evaluation flows.

Harsh’s take

The day showed clear movement toward longer context and longer runtime rather than raw benchmark gains. GLM 5.2 and the sandbox extension both target the same pain point of agents that lose state mid-task. At the same time, consumer surveys indicate that simply labeling a feature AI can reduce trust, which undercuts many of the new device integrations.

Second-order risk is that teams will wire these longer sessions into production before they add proper evaluation harnesses. The LangChain posts on loop engineering and human judgment loops arrived at the right moment but will be ignored by builders chasing the new context window.

Run one controlled test this week that compares GLM 5.2 against your current model on a 50-step agent task while logging every retry and cost spike.

by Harsh Desai

Sources

Vendor launches

Replicate new models

Hugging Face trending

Product Hunt picks

Other

Industry news

More AI news

Everything AI. One email.
Every Monday.

New tools. Model launches. Plugins. Repos. Tactics. The moves the sharpest builders are making right now, before everyone else.

No spam. Unsubscribe anytime.