If you like building agents that get work done, you're in for a treat in this newsletter. CUGA is a new agent framework that is topping benchmarks and using all sorts of cunning under the hood to help you build better agents that can execute complex tasks.

There is also news on model releases, code execution sandboxes, and the latest podcast episode from The Flow, all on OAuth and MCP.

Phil Nash​
Developer relations engineer for Langflow

🛠️ Building with AI, Agents & MCP

IBM Research releases CUGA

CUGA is IBM's ConfigUrable Generalist Agent, an open-source agent framework that you can use to build agents that are good at complex task execution, can use MCP servers, OpenAPI specced API, and custom tools, and has configurable reasoning modes. It topped the AppWorld benchmark and is in second place on the WebArena benchmark, and it has been integrated in Langflow (though that's not been released publicly yet, you can sign up to hear more or build the latest from the main branch). Check out the source code and how to run CUGA directly on GitHub.

Agent builders

There have been a lot of releases of various agent builders, so before you dig into the following news, check out this guide on choosing an AI agent framework in 2025.

Now for the releases: ElevenLabs released Agent Workflows, and Microsoft combined Autogen and Semantic Kernel into the Microsoft Agent Framework. We also compared building a simple agent with Langflow and OpenAI's AgentKit.

New model alerts 🚨

Anthropic released Claude Haiku 4.5 which appears to be faster, cheaper and, on several benchmarks, better than Sonnet 4. Meanwhile, Microsoft AI released their first image model, and Google Deepmind trained a specialized version of Gemini 2.5 Pro on computer use and added grounding via Google Maps.

Claude gets some new skills

Anthropic announced that Claude now supports Skills, where Skills are a collection of markdown content and scripts that the agent can run. It seems to me surprisingly close to MCP, but it is intended to mitigate MCP tools filling the context (something Pulse MCP tried to fix recently with agentic MCP configuration). I enjoyed Simon Willison's take on Skills as well as the Anthropic engineering team's deep dive into creating the feature.

Building with in-browser AI

This in-depth blog post goes into how to build the classic game Guess Who? in the browser using Chrome's built-in Prompt API.

Cloudflare releases Sandboxes

Getting LLMs to generate code is second nature to developers at this point, but if it's part of an application then having a safe sandbox in which to run the untrusted code is important. Sandboxes is Cloudflare's isolated environment to help you build this safely.

How to Use OAuth in MCP with Max Gerber from Stytch

In the latest episode of The Flow we bring on Stytch's Max Gerber for a hands-on tour of how OAuth powers secure AI agents and MCP servers, complete with live demos and practical tips on making your app an OAuth provider so users can safely bring their own agent.

🗞️ Other news

• The State of AI report 2025 was released. It's a big report, be warned!
• Andrew Ng released a new DeepLearningAI course on Agentic AI
• Anil Dash opines that the majority AI view in the tech industry is that "technologies like LLMs have utility, but the absurd way they've been over-hyped, the fact they're being forced on everyone, and the insistence on ignoring the many valid critiques about them make it very difficult to focus on legitimate uses where they might add value." I hope you're out there adding value with the agents you build.
• Phil Schmid describes the shift from shallow loops to deep agents
• An Anthropic study showed that you can produce a "backdoor" vulnerability in an LLM with just 250 malicious documents added to the training data.
• This is a great recap of a panel on context engineering. We're well beyond prompt engineering now.
• This quick post on building production RAG systems gets straight to the lessons

🧑‍💻 Code & Libraries

• ElevenLabs released a number of open-source agent and audio components
• A list of great resources for building with Langflow: https://github.com/GovindMalviya/awesome-langflow
• An open, lightweight, event-based protocol for AI agents to connect to user-facing applications: https://github.com/ag-ui-protocol/ag-ui
• An MCP server that uses Redis to give persistent memory to your agents: https://github.com/joseairosa/recall
• Give your agents the power of generating videos with this unofficial Sora MCP server: https://github.com/Doriandarko/sora-mcp
• A new SDK for embedding interactive UIs in chats apps, including ChatGPT Apps and MCP-UI: https://github.com/fractal-mcp/sdk

🔦 Langflow Spotlight

You know that you can build MCP servers in Langflow, but with the release of Langflow 1.6 using MCP servers got an upgrade. The sidebar now stores all the MCP servers you have configured, so you can grab them and use them in any of your flows. There is also a dedicated area to manage your MCP servers. Giving your agents access to great tools has never been easier!

🗓️ Events

The Langflow team are out east in Australia and Malaysia in the next couple of weeks.

October 25th, Brisbane - Catch me at the GDG Devfest in Brisbane talking about AI on the web.

October 29th, Melbourne - Me again, this time talking web AI at the Melbourne GDG meetup.

November 4th, Kuala Lumpur - You can find Tejas speaking about thriving as a professional, particularly leveraging AI, at Infobip Shift Kuala Lumpur.