Agents are starting to take on a life of their own, and as agent builders we need to consider the potential outcomes. The story of the autonomous agent that was denied when it opened a pull request to contribute to matplotlib and consequently wrote a hit piece on the maintainer had opinions on the internet veering between a fabrication or the beginning of Skynet. I think the lesson we should be taking away from this is that the the end user of an agent is not the only human that may come into contact with it and we should design accordingly.

In this edition we have practical ways to build with subagents, build voice agents, and build memory for agents. Plus plenty of model releases and a reminder not to get LLMs to generate passwords for you.

​Phil Nash​
Developer relations engineer for Langflow​

🛠️ Building with AI, Agents & MCP

Context control

Taking control of your agent's context is key to their success, this article on how LangChain improved their coding agent by improving the harness has some good lessons on how to take advantage of tracing and self-verification. On the other hand, you might be trying to keep your token costs, and thus the size of your context, down. This article on the LLM Context Tax has actionable ways to use caches and avoid sending excess context. Here's an in-depth look at how prompt caching works within OpenAI.

Different ways of building agents

This developer built a sub-500ms latency voice agent from scratch and this is a good practical project on how to build memory for your agent with Python and SQLite. If you want to go beyond agents to subagents, this explainer from Builder covers a lot of ground.

As agents get more complex so do evals. Finding out what to eval, and that sometimes the thing you built didn't help, were part of this developers first evals.

Building skills

This is a great guide on how to build an Agent Skill for a developer tool. Did you know that Skills are supported in the OpenAI API now too?

🧠 New models

It's been a little while since I sent a newsletter, so there has been a lot of new models. Some interesting ones are:

• Anthropic's Claude Opus 4.6 which now comes with a 1 million token context window and a "fast mode", as well as Sonnet 4.6​
• For coding, OpenAI released GPT-5.3-Codex and the smaller but faster GPT-5.3-Codex-Spark. In the general purpose line, GPT-5.3 Instant is an update focusing on tone, relevance and flow
• In open models, the team at Alibaba released Qwen Image 2 and the multimodal Qwen 3.5, with many sizes that can be run locally
• Google was busy, rolling out Gemini 3.1 Pro, Nano Banana 2, Gemini 3.1 Flash-Lite and Lyria 3 for creating music​
• And in embedding model news, Perplexity released some new models based on Qwen3 models, but outperforming them across a number of benchmarks

🗞️ Other news

• The latest edition of The Flow covers the rise of MCP Apps as David and Roy talk with Ruben Casas from Postman
• ​WebMCP is a proposal that allows web developers to expose web application functionality as tools to AI agents or browser assistants that are using their applications. Check out how it works on GitHub​
• The email platform Resend wrote about how to provide your agent, OpenClaw in this case, with an email account. But remember, you're responsible for who they decide to send emails to!
• A piece of research showed that for non-reasoning LLMs, repeating your prompt improved the outcome. So much for trying to save tokens!
• In a good reminder that LLMs are probabilistic generators, don’t use them to generate passwords​
• The OpenAI Responses API now supports websocket connections for long-running, tool-calling workflows

🧑‍💻 Code & Libraries

• ​Hugging Face‘s Transformers.js v4 preview is out. It can run AI models in the browser, Node.js, Bun, Deno, basically wherever JS runs
• Hugging Face have also added support to Transformers for mixture of experts models​
• ​llmfit is a terminal tool that finds models that will run on your machine
• ​Docker Model Runner brings the very popular inference engine vLLM to macOS
• ​MachineAuth is authentication and permission infrastructure for agents to access services
• ​zvec is an in-process vector database
• ​helm is a TypeScript framework for building agents with skills and permissions
• ​Utah is an agent harness built using durable execution powered by Inngest. Read more about it in this blog post​

📅 Events

I'll be out at OpenSearchCon China on March 16th-17th, speaking about document ingestion with Docling and OpenSearch with Carol Chen from Red Hat.