Agent SDKs are evolving

Is a token crunch coming? This week GitHub paused sign-us for GitHub Copilot Pro, Pro+ and Student plans, tightened up their usage limits, and removed Opus from their Pro plans. And today, Anthropic seemed to remove Claude Code from new Pro plans, though that has been reversed quickly. In general, while this is only seeming to affect individual plans related to coding agents, it could point to an inflection point where AI companies start considering how their pricing matches up to their costs.

In other, more amusing, news the shoe manufacturer Allbirds is pivoting to AI compute infrastructure. Selling inference is hot, I just never expected a move from shoes to GPUs.

Phil Nash Developer relations engineer for IBM

🛠️ Building with AI, Agents & MCP

AI engineer Europe videos

AI Engineer Europe was a couple of weeks ago and the videos are coming out in this YouTube playlist. There's plenty of content on Coding Agents, OpenClaw, MCP and more. One of my takeaways was on how split the community was on speeding up vs slowing down when writing code with Ryan Lopopolo of OpenAI who banned his team from touching their editors going up against Mario Zechner, creator of the coding agent pi, saying "Slow the **** down". You can also catch my talk on an open-source stack for RAG.

Langflow 1.9 released

The latest version of Langflow is now live, with version 1.9 bringing features like an assistant to help you build custom components, the Flow DevOps Toolkit to help you version, test and deploy flows, and support for connecting Langflow itself to your coding agents via MCP.

What's the future of MCP?

In more video news, the keynote from the MCP Dev Summit has been published. It covers where MCP is going in 2026 including features like the stateless redesign, long running tasks, and skills over MCP.

Cloudflare shared their reference architecture for MCP deployments, covering a lot of detail particularly in the security aspects.

Evolving agent SDKs

Both OpenAI and Cloudflare announced evolutions of their agent SDKs both highlighting how they will help build harnesses for longer running agents. Sandboxes and durable execution both appear as features for each SDK and show where these companies think agent style execution is heading.

On the server side, Google introduced Deep Research Max, an API based agent with comprehensive research capabilities.

🧠 New models

🗞️ Other news

• ALTK-evolve is a memory system for agents that helps them learn and use guidelines for decision making from previous executions
• IBM research released VAKRA, a new benchmark for evaluating how well AI agents reason and act in enterprise-like environments
• KellyBench is another benchmark, this one focused on how agents perform in sports betting markets. I'm not sure we should be optimising for gambling, but at the moment they all lose money.
• I still prefer MCP over Skills
• The Slack Engineering team share how to manage context in long-run agentic applications
• Using too many tokens? Try caveman mode!

🧑💻 Code & Libraries

• Check out all the AI examples in this awesome-ai-apps repo
• Manifest is an open-source smart model router
• Chunky allows you to convert and validate document parsing and chunking
• OpenWebUI released a desktop app

🔦 Langflow Spotlight

One of the biggest additions to Langflow 1.9 is the Langflow Assistant. It generates and updates custom components for you directly within the Playground.

This makes it even easier to build out your own components to complete your flows.

🗓️ Events

This week you can find Phil at NDC Sydney talking about how to build an agent harness.