UCP Updates: Pass Latest Spec and Optimize Agent Browsing

Google just posted v2026-04-08 of the UCP spec — the standard that supports full-on commerce inside of Gemini and AI Mode. It's how you get listed... 𝘢𝘳𝘦 𝘺𝘰𝘶 𝘳𝘦𝘢𝘥𝘺? Most are still figuring out where to start. 𝗥𝘂𝗻𝘁𝘆𝗽𝗲 𝗷𝘂𝘀𝘁 𝗯𝘂𝗶𝗹𝘁 𝘁𝗵𝗲 𝗮𝗻𝘀𝘄𝗲𝗿. Today we're launching full end-to-end UCP support in Runtype — 𝘥𝘦𝘷𝘦𝘭𝘰𝘱, 𝘷𝘢𝘭𝘪𝘥𝘢𝘵𝘦, 𝘢𝘯𝘥 𝘴𝘩𝘪𝘱 𝘜𝘊𝘗-𝘤𝘰𝘮𝘱𝘭𝘪𝘢𝘯𝘵 𝘈𝘐 𝘱𝘳𝘰𝘥𝘶𝘤𝘵𝘴, 𝘧𝘢𝘴𝘵𝘦𝘳 𝘵𝘩𝘢𝘯 𝘢𝘯𝘺𝘰𝘯𝘦 𝘦𝘭𝘴𝘦. For platforms: get the tools to ensure full spec compliance. For builders: a production-ready AEO product template with UCP baked in, plus everything you need to tune your own agents. Make them run 3x faster at half the cost. For everyone: we're open sourcing the demo UCP server running at ucp.runtype.dev. Build fully featured UCP clients like we did. Watch the video for the deep dive! Then validate how your platforms and agents are working with UCP in minutes at runtype.com

🚨 I found #AISupplyChain attack path where poisoned RAG/web content can silently influence an MCP enabled AI assistant to connect to an attacker controlled server and execute remote code on the backend. The attack surface: untrusted documents + dynamic tool onboarding. The gap: existing defenses don’t treat retrieved web content as part of the tool supply chain. The impact: a single web page can become the entry point to RCE in the assistant’s environment. I’ve published the research write-up and demo on GitHub : https://lnkd.in/gve7KczQ #AISecurity #LLMSecurity #SupplyChainSecurity #MCP #PromptInjection #RAG

NEW paper from Alibaba. A 30B MoE with only 3B active params matches Qwen3-235B on real tool-use workloads. AgenticQwen-30B-A3B: 50.2 average on TAU-2 + BFCL-V4 Multi-Turn. AgenticQwen-8B: 47.4. Both more than double their vanilla Qwen baselines and close most of the gap to a 235B model. How: two RL flywheels run in parallel. - The reasoning loop mines the model's own errors into harder problems each round. - The agentic loop grows simple linear tool-use trajectories into multi-branch behavior trees. - Simulated users actively try to mislead the agent. The training distribution gets harder on its own. Why it matters for agent devs: you can stop paying frontier prices for routine tool-use workloads. And the flywheel recipe is reusable. Generate your hard examples from your own agent's failures, not from static synthetic data.

I've been experimenting with this personally for quite some time, and today I'm excited to finally make it public. Introducing OCP — Open Context Protocol (v0.1.0) 🚀 https://lnkd.in/dMxpuT8P OCP is an open context layer built on top of MCP that adds persistent, retrievable context and coordination primitives for AI agent An open layer built on top of MCP that helps AI agents remember context across sessions. Website: https://lnkd.in/dMxpuT8P GitHub: https://lnkd.in/dg5SAezv PyPI: https://lnkd.in/dqPXmJPW Licensed under Apache-2.0 (code) and CC-BY 4.0 (specification). This is still an early v0.1.0 release, the foundations are there, but there's a huge amount of room to grow. I'd genuinely love feedback, ideas, and contributions from the community, whether that's embedding backends, storage adapters, integrations, spec discussions, or simply testing it and opening issues. If you've shipped agent memory in production: what's the one piece of context your agents keep losing? https://lnkd.in/dMxpuT8P #OpenSource #MCP #AIAgents #LLM #DeveloperTools #OpenStandards #OpenContextProtocol

Most developers use Claude like a slightly smarter Google. Here's the difference between using it and actually leveraging it. 7 things I wish someone had told me in week one: → System prompts aren't optional — they set the entire operating mode → Context window ≠ memory. Manage it intentionally → Tell Claude *who it is*, not just *what to do* → Multi-turn chains beat single mega-prompts → Use XML tags for structured output — it's dramatically cleaner → Ask it to think step-by-step before answering complex problems → Test adversarially — see where it breaks before your users do. Saved me months of trial and error. Follow for more. #ClaudeAI #AITools #DevTips #AIForDevelopers

We recently made some upgrades to the Unstructured docs to make them more useful to both humans and agents 👇 We’ve restructured all API operation information into a concise collection of operation-level reference topics. The API Reference now comes complete with real-world code samples, and an interactive playground that lets you exercise APIs and see actual calls and responses, right from within each topic: https://lnkd.in/euF_e6bS We’ve also separated out our workflow nodes content into a new Workflow nodes section that presents node information in a concise, sample-first reference: https://lnkd.in/emVh4Ghx And we’ve updated and centralized our list of Available models per node type: https://lnkd.in/ess3vSsk For agents, this means a much more concise llms.txt and llms-full.txt that’s easier to parse and consume. Learn more here 👉 https://lnkd.in/exZNs2XQ

Codex for (almost) everything: Codex can now use apps on your Mac, connect to more of your tools, create images, learn from previous actions, remember how you like to work, and take on ongoing and repeatable tasks. 🔗 OpenAI #Codex #AItools

Deploy and manage your Insites instance directly from Claude Desktop or VS Code Introducing CloudShell — a new MCP server that connects Claude Desktop and VS Code to your live Insites instances over HTTP, with no CLI installation or browser login required 🚀 Sync files, batch deploy, pull logs in real time, manage environment variables, and list installed modules through a single conversational interface using the AI tools you already use. Read more about CloudShell here: https://lnkd.in/giJzttWf #CloudShell #AI #DeveloperTools #VSCode #Automation

We shipped a LlamaParse MCP server to let you parse, classify, split, and generally operate over your hardest documents with your favorite AI agent 📄🤖 Check out the MCP server: https://lnkd.in/gkAQGFjM This is both a useful feature and a great piece of engineering that we want to share with the community. 💡 MCP does not have built-in file upload support, so we needed to implement a URL-based upload endpoint and couple it with parse operations 💡We built in an integration with WorkOS OAuth. 💡We built in observability and rate-limiting Huge shoutout to Clelia Astra Bertelli for shipping this! Come check out our blog writeup: https://lnkd.in/gGzbDMFW LlamaParse: https://lnkd.in/gF2VHrQk

Your MCP server code passes every test, but the agent still fails. Why? Traditional testing wasn't built for the non-deterministic nature of LLMs. As Red Hat engineers Wesley Chun and Calum Murray shared at the MCP Dev Summit, there is a massive "testing gap" between server code and agent behavior. The solution? MCP Checker. This open-source framework verifies that models actually understand and use your tool semantics—catching silent failures and dangerous "workarounds" before they hit production. Read the full deep dive on the Agentic AI Foundation blog: https://bit.ly/4u2ikug