The AI Engineer World’s Fair was in San Francisco last week. No, I wasn’t there, and yes, I was feeling all of the FOMO. If you’re in the same position, then you can catch up with what went on with Latent Space’s write-ups from the event and subscribe to the YouTube playlist as it starts to to publish the talks. You can also check out the talks from the online track while you wait.

In this edition of AI++ we take a look at the skills it takes to build great Skills, get hands-on with some agent harnesses, and find out what happened when a developer invited 2,000 people to hack their AI assistant.

​Phil Nash​

Developer relations engineer for ​IBM

🛠️ Building with AI, Agents & MCP

Fable returns

Anthropic and the US government have sorted out their issues (until the next jailbreak, I guess) and access to Fable 5 and Mythos 5 (for those in Project Glasswing) has been restored! If you’re looking to work better with Fable 5, check out this post or talk by Thariq Shihipar from the Claude Code team.

OpenAI fans should watch out on Thursday as GPT-5.6 Sol, Luna and Terra will be launching.

Harnesses

Two teams showed practical examples of working agents using their harnesses that you can get inspiration from and build upon. First, the IBM Research team wrote about how to build real agentic apps with CUGA and the team at Coral showed how they adapted their Reef harness to beat benchmarks.

For an in-depth look into the theory of harnesses check out Lilian Weng’s Harness Engineering for Self-Improvement.

Skills

Matt Pocock presented on the AI Engineer World’s Fair online stage on building great Agent Skills, my favorite tip was that you can hide Skills from the agent unless you invoke them directly. Laurie Voss also shared tips on writing good Skills with a couple touching on testing and evaluation. On theme with this is Microsoft’s SkillOpt that can help you benchmark and train your Skills.

While not directly Skill focused, the VS Code team wrote about how they optimized the system prompt in GitHub Copilot.

🧠 New models

• As mentioned up top, OpenAI will be releasing GPT 5.6 this week. If you want to know what’s coming, check the preview of GPT-5.6 Sol and system card for GPT-5.6​
• Amidst all the excitement around Fable, we can’t ignore that Anthropic also released Claude Sonnet 5​
• On the other side of the scale equation, LFM2.5-230M is a small, 250m parameter model that is good at tool calling. It’s not to be used for reasoning heavy workloads, but is good at skill selection, and is built for on-device agentic workloads
• In the world of image and video generations, Google released Nano Banana 2 Lite and Gemini Omni Flash​

🗞️ Other news

• ​What happened after 2,000 people tried to hack my AI assistant​
• ​Prompt Injection as Role Confusion​
• ​"It's Hard to Eval" Is a Product Smell​
• Multi-model harnesses have blown up recently, and this week the team at Cognition shared their work on Devin Fusion. It uses Fable, along with a smaller sidekick model that performs smaller tasks, while the big model keeps control of orchestration. It turned out both cheaper and better than Fable on its own
• Gergely Orosz noticed this too and wrote about a new trend, smart model routing​
• ​Sandboxing an AI Agent​
• ​Designing for GenUI: What Changes When AI Builds the UI?​
• ​Autoresearch, Claude and Constrained Optimization​
• ​How we taught a small LLM to throw away 68% of our RAG context​

🧑‍💻 Code & Libraries

• ​agentos is a lightweight VM that you can use to isolate agent written code within your app
• Vercel released AI SDK version 7 and as part of it you can now control other harnesses through the SDK​
• ​Persona is a vanilla JS, themable, pluggable AI chat widget
• ​fenic is a Python library for Semantic DataFrames that work for both humans and agents​
• ​OpenWiki is a CLI that writes and maintains agent documentation for your code base​
• Valmis is an AI Agent designed for work with security in mind​

🔦 Langflow Spotlight

Agents often need to work with large collections of files, but loading every file directly into the prompt does not scale well. These File Processing components help separate file discovery from file retrieval inside Langflow workflows.

• The File Description Generator creates searchable natural-language descriptions for each file
• The File Content Retriever exposes tools that allow an agent to retrieve specific file content by path

When you’re smart about your context like this, your agents will thank you.

🗓️ Events

On the 9th July, David and Tejas are using IBM Bob to build a NotebookLM Clone powered by OpenRAG. Check in on their past streams to see what they’ve been up to and join them this week to see how they get on.

If you’re in the Boston area on 28th July, head along to Hacking Agents to learn more about OpenRAG, Unstructured, IBM Bob and PyData. And if you can spend the whole day, check out the Agentic AI Coding workshop.