AI++ // MCP tools are for the LLMs and how to stop models scheming


Over the last weekend the Langflow team was out at the CascadiaJS conference and Cascadia AI Hackathon. It was inspiring to meet so many developers, work with them on their AI hacks, and see fewer demo hiccups than a Meta product launch. The winning team built a full music sequencer, MIDI keyboard and visualizer that generated beats that could then be edited by hand or further with AI. Congratulations to the winning teams and to everyone who built something and learned something new over the weekend (and πŸ‘‹ if I met you at any of the Cascadia events).

This week in AI++ more people share how they build AI agents, there's an open source clone of a popular agent, and we reconsider the best ways to write MCP tools. Plus: how to stop agents scheming, how NotebookLM was designed, and a cursed programming language.

Phil Nash
Developer relations engineer for Langflow

πŸ› οΈ Building with AI, Agents & MCP

​Engineering an AI app

From interaction models to reasoning strategies, this article goes through many of the choices you have to make when building an AI app. It's all based on the experience of writing apps like new.email.

πŸ“Ί FAANG PM reveals how to build AI agents​

This is a great interview/live coding/vibe coding session where a PM veteran shows how you can build AI agents that really work, featuring Langflow for back-end work and v0 for the front-end.

​Build MCPs for LLMs, not developers​

This article from Vercel looks into how to better design MCP servers to optimize for LLM memory and context. It's time to unlearn everything you thought you knew about API design to move into effective MCP design.

The Anthropic team also shared a post on writing effective tools for agents with a lot of detail.

​Chunking strategies to improve your RAG performance​

This article from Weaviate probably goes deeper than you ever thought you'd need to about chunking data. But when you're building RAG-powered AI apps that rely on the right context at the right time, the way you chunk matters.

Before you can chunk you need to extract the data, and IBM just announced Granite-Docling, a tiny model for document understanding. Docling is a powerful way to extract data and is available inside Langflow too.

​OpenPoke: recreating Poke's architecture​

Poke is an agent you can chat to in iMessage. Not satisfied with just using Poke, this developer wrote about reverse engineering a bunch of its features to release an open-source version called OpenPoke.

πŸ—žοΈ Other news

πŸ§‘β€πŸ’» Code & Libraries

πŸ”¦ Langflow Spotlight

If you're building agents in Langflow you know that you need to hook up tools to your agent. The simple agent template shows you two such tools, the calculator and URL. But did you know that you can turn almost any Langflow component (including the agent component itself) into a tool, just by flipping a switch.

It's called Tool Mode and it enables all the tools a component can provide, allowing you to edit their names and descriptions. So, if you want to make search APIs, data ingestion or other agents available to your agent, just toggle Tool Mode.

​

πŸ—“οΈ Events

October 8th, 11:00 AM PDT – Security is the topic of The Flow as David and Carter chat with Max Gerber from Stytch. Catch it live or subscribe to the podcast to get the recording.

October 15th–16th – join the Langflow team at AI for the rest of us in London. Promising two days of inspiring and practical sessions that demystify jargon, share real world stories of AI in action, and include developer focused sessions and hands-on workshops, this will be a fantastic community event. You can save 20% off the cost of a ticket with the code LANGFLOWCREW.

Enjoy this newsletter? Forward it to a friend.

2755 Augustine Dr, 8th Floor, Santa Clara, CA 95054
​Unsubscribe Β· Preferences​

AI++ newsletter

Subscribe for all the latest news for developers on AI, Agents and MCP curated by the Langflow team.

Read more from AI++ newsletter

Who is naming these model releases? Gemini’s new image editing model is called Nano Banana (and they dropped a great tutorial on how to use it as a developer). Meanwhile, Microsoft AI launched two models in the new MAI series, presumably pronounced β€œmy” because they’re not OpenAI models. At least OpenAI, the creators of gpt-4o-mini and o4-mini, were a bit more sensible with their recent release of gpt-realtime along with the generally available Realtime API. And that’s only some of the newly...

Privacy has been a hot topic for consumer AI applications this month. Users have been finding out that chats with OpenAI, Grok, and MetaAI that they shared, were being indexed by Google and showing up in search results. Sharing a chat with your friend doesn't mean you want to share it with every search engine out there, so make sure your users know what they're doing and don't forget your robots.txt. Multi-agent systems are the theme for this newsletter, with tutorials on how to build them...

OpenAI dominated the AI headlines last week with the release of new open weights models (their first open release since GPT-2) and GPT-5. There’s been some user backlash against GPT-5 replacing all the models in ChatGPT, but as developers we can still pick the GPT we want to use. Not wanting to feel left out, Anthropic released Claude Opus 4.1 and the Qwen team impressed with Qwen-Image. This week in AI++ we have some great open-source examples of agents, we dive into UX and accessibility for...