AI++ // are MCP servers inefficient? What can we learn from coding agents?

There has been a flurry of new frontier models dropping over the last week that you can already use in your applications. Gemini 3 was released today, and Grok 4.1 and GPT-5.1 both arrived last week.

This week we're also learning a lot of lessons from how coding agents are built, including building a coding agent in Langflow if you want hands-on experience with your own. There's also much debate over the efficiency of MCP and whether other tools fit the job better.

Phil Nash
Developer relations engineer for Langflow

🛠️ Building with AI, Agents & MCP

New model releases ✨

You know what it's like, you're trying to get on with building your world changing agent, and then three new models arrive on the scene within a week. Gemini 3 has just launched and it's has topped the LMArena leaderboard. Gemini also released some interesting new features, including a file search tool and improved structured outputs.

Grok 4.1 arrived just a couple of days ago, topping the LMArena board only to be unseated by Gemini. Still impressive.

It almost seems old news that GPT-5.1 was released a week ago and shortly after in the API, but things move fast in AI!

Lessons from coding agents

Coding agents have been one of the breakout successes so far, so there must be much to learn from them. AmpCode wrote a guide on context management, including some of the features they built into their agent that allow for intentionally including data in the context, as well as editing, restoring and forking context.

Cursor wrote about improving their coding agent with semantic search. The post also highlights using offline and online evals to recognise that a change improves the agent.

If we can learn about all agents by looking at what they can do now, here's a view into the future of agentic coding to take some inspiration from.

And again, you can start your own coding agent from the comfort of Langflow. See how good you can make with just drag-and-drop.

Optimizing MCP

There is great debate over whether MCP is inefficient and whether you would be better having your agent execute code or use a CLI. Perhaps some MCP servers are poorly implemented and using a CLI will result in poorly implementing MCP anyway?

With that in mind, Code-Mode is a library intended to help your agents call tools by executing code and MCP-Optimizer is an intermediary MCP server that optimizes the tools available to an agent.

🗞️ Other news

Rachel Lee Nabors wrote a couple of articles for web.dev on why to pick the right size of model for your feature and how to choose small, efficient models
Google have released a paper on context engineering covering sessions and memory in great detail
Can you optimize your prompts by saying less?
Two resources for considering the UI and UX of AI: The Shape of AI and UI for AI
Here's some inspiration from Anthropic of what you can build with Claude

🧑‍💻 Code & Libraries

VS Code’s GitHub Copilot extension is becoming more and more open-source
Evalite is a new evaluation framework written in TypeScript
Google released their ADK toolkit for Go
Stripe have shared their toolkit for AI and Stripe
If you build for mobile in Flutter, then there’s now a library to introduce generative UI to your apps
Mozilla launched any-llm, a single interface to interact with any LLM provider, remote or local

🔦 Langflow Spotlight

I normally like to show off a feature or component of Langflow in this section, but this week I want to share a blog post from Luiz Henrique Salazar on how to deploy Langflow with custom components on Kubernetes. It’s a fantastic and in-depth tutorial that walks you through a lot, from deploying Langflow with a Helm chart, to building a custom component and deploying it with just the back-end Langflow runtime.

For a simpler deployment recipe, check out Tejas’s post on deploying Langflow to services like AWS, Fly.io, Render, and Hetzner.