Happy 2026! I hope your year has kicked off well and that you're looking forward to a year of building ever more impressive and successful agents. As we head into 2026 I'm looking forward to seeing how MCP continues to evolve under the Agentic AI Foundation, the blossoming of many AI-powered workflows, and how harnesses grow to help us build even better agents. What do you think will be the growth areas for the next few months?

In this edition of AI++ we check in with some reviews of 2025, consider how less can be more for some agents, and take a look at memory and context.

​Phil Nash​
Developer relations engineer for Langflow​

🛠️ Building with AI, Agents & MCP

A real life agent

​Reachy Mini is a programmable desktop robot from Hugging Face. This walkthrough shows how you can turn Reachy into an agent using locally hosted NVIDIA models. Don't worry if you don't have a robot of your own, there is also a simulator that you can play with. This is a fun, and expressive, way to interact with your agent.

Langflow 1.7 released

Just in case you missed it at the end of last year, Langflow 1.7 was released. The new features include Streamable HTTP support for MCP clients and servers, new components for flow control, and a powerful new agent component CUGA.

Fewer tools makes for a better agent

Vercel have an agent called d0 that they use for understanding their data. After spending a long time building a complicated system with many specialized tools, they stripped a lot away from their agent, reducing the number of tools and ways they tried to control the model, and found that it worked better. As models improve, getting out of the model's way can result in better outcomes.

This is a similar story of a developer that wrote an agent that just had access to bash and some custom scripts. The agent wasn't over-engineered but could easily control a web browser for their dev tools use-case.

What is memory?

"If you're building agents without learned memory, you're leaving performance on the table." is the bold claim from this article on how memory makes for a good agent.

Speaking of memory, Factory looked into evaluating ways to compress memory. Retaining memory while avoiding context rot is going to be key for long running agents this year.

🗞️ Other news

• Andrej Karpathy reflects on LLMs in 2025, from the emergence of reinforcement learning from verifiable rewards to coining the now ubiquitous phrase "vibe coding"
• Simon Willison wrote a much longer recap of 2025 covering prompt-based image editing, long tasks, vibe coding, MCP and slop
• AI browsers are one of the riskiest ways to use AI today, but OpenAI have published some interesting research on how they are hardening Atlas against prompt injection.
• ​MCP is exploring moving to become a stateless protocol, which should help scaling

🧑‍💻 Code & Libraries

• Version 6 of Vercel's AI SDK was released: https://vercel.com/blog/ai-sdk-6​
• Tambo is a generative UI SDK for React: https://github.com/tambo-ai/tambo​
• This project highlights which of the MCP spec's capabilities are supported by various clients: https://github.com/apify/mcp-client-capabilities​
• A collection of agent skills for context engineering: https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering​
• Open Tinker is open source infrastructure for reinforcement learning: https://github.com/open-tinker/OpenTinker​
• There are a lot of awesome agentic patterns here: https://github.com/nibzard/awesome-agentic-patterns​

🔦 Langflow Spotlight

In Langflow 1.7 there are new components that help you control a flow dynamically. One of those components is the Smart Router. It's an LLM-powered variation of the If-Else component, but it uses a model to judge an input by a criteria that you provide and then creates ports for different routes out of the component. This is a far more powerful way to control a flow and means your flow can be intelligently routed based on any content that flows through it.

📅 Events

On January 29th IBM is hosting AI Demystified, a virtual event and hackathon covering many different open-source AI tools including Langflow, CUGA, Granite models, and the Agent Lifecycle Toolkit (ALTK).