The last couple of weeks has seen students booing commencement speakers at graduation ceremonies in Florida and Arizona when they mentioned AI. This is a visceral reaction to what they see as a threat to their careers. Meanwhile, developers working with AI are burning out and getting “Brain Fry” from doing more work at higher intensities without the same fulfillment. I wrote a bit about this myself, sharing that I found it hard to be proud of a useful little app that I built. AI is changing work, and we all need to work together to understand it, make it sustainable, and ensure we are bringing graduates on board so that we can all learn how to work in the age of AI together.
With that said, the AI industry forges ahead at the usual blistering pace. In this edition of AI++ we take a deep dive into Agent Harnesses, review the launches from Google I/O and check in on MCP. We also learn what happens when you put agents in charge of their own radio station.
Phil Nash
Developer relations engineer for IBM
🛠️ Building with AI, Agents & MCP
Agent Harnesses
2026 is being dominated by talk of how the agent harness is as important, if not more, than the model driving it. How we speak about AI engineering changes often, so if terms like scaffolding, harness and agent all sound like the same thing, this glossary of agent terms from Hugging Face is a good start. If you want to see what goes into a harness and how it tethers a model to the behaviour you want then check out Tejas’s talk from the AI Engineer: Europe event and check out this course on learning harness engineering.
In the world of Langflow, there is a new Policies component that helps you build out the harness for your agent by constraining tool use with natural language policies. You can describe what an agent should and shouldn’t be able to call and Langflow will turn the policies into deterministic rules.
Other ways to build out your harness include taking advantage of agent hooks to control the workflow, or wrapping your agent in a firewall like Claw Patrol. Sandboxes are also important for running agent generated code, CoreWeave released their Sandboxes product, but if you want to run your own, the Kubernetes project built agent-sandbox for you to deploy yourself.
Google I/O
Last week was Google I/O, which came with at least 100 new announcements. Things to keep an eye on include
- Gemini Omni: a new model that can ingest text, images, video and audio and output anything, although currently you can only use it to generate video. API access, and hopefully other output modalities, are coming soon
- Gemini 3.5 Flash out performs Gemini Pro 3.1 and is generally available. The developer guide is worth a read for getting the most out of it. Some people have been reporting that it uses more tokens, so ends up costing more. If you are finding this, try the Low thinking level, which outperforms Gemini 3 Flash High for 45% fewer tokens than Flash 3.5 Medium
- Antigravity SDK: like all the other coding agents, you can now use the Antigravity SDK to build your own agent
- Agent Executor: a new, open-source, distributed agent runtime that provides durable execution, secure isolation, and other features for running your agents at scale
MCP
The MCP project is preparing to release a new version of the specification which includes changing from a stateful protocol to stateless. This is going to help scale MCP servers so is worth reading up on. MCP Extensions are also becoming a first-class feature of the spec and there are two official extensions, MCP Apps and Tasks. There’s much more so check out the release candidate.
In the world of the browser, WebMCP is going to origin trial in Chrome 149. This means you can start testing it out with real users. WebMCP is not exactly the same as MCP, it is intended to allow you to expose tools to agents that can browse the web, so agents can complete tasks on website with greater efficiency.
🧠 New models
- The Qwen team released the proprietary Qwen 3.7-Max which you can use through the API. Look out for smaller, open models in the 3.7 series!
- Cohere released the open-source Command A+
- PrismML released a new image generation model called Bonsai, based on FLUX.2 Klein 4B. It’s a 4B model that uses binary or ternary weights, which dramatically reduces the size and means it can run on an iPhone, or even in the browser. It’s a bit of a step backwards compared to frontier image generation models, but the fact that you can run it on a mobile device and get some good results is a step forward
- In the world of retrieval, the Ettin Reranker Family are cross encoder models that have great results, even for the smallest of the family with only 17 million parameters
- And if you’re looking for a multilingual embedding model, IBM’s Granite Embedding Multilingual R2 might be the model you’re looking for
🗞️ Other news
🧑💻 Code & Libraries
🔦 Langflow Spotlight
Langflow flows can run code using the Python Interpreter component. You can even provide the component to an Agent component to allow the agent to execute Python code, giving it predefined imports. Use it for tasks like:
- Working with data using packages like pandas
- Running calculations that need Python logic
- Creating small scripts inside an agent workflow
- Processing structured data before returning an answer
🗓️ Events
Catch the live broadcast of The Flow at 1pm EDT on 27th May, where hosts David and Mike will be speaking with Prathmesh Patel, CEO of MCPJam, about token bloat.
On June 3rd and 4th, I will be speaking at NDC Copenhagen about building agents and building UI for agents with MCP Apps. Come say hi if you’re going to be there!