𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗻𝗲𝗲𝗱 𝗹𝗮𝗽𝘁𝗼𝗽𝘀 - 𝗹𝗶𝗸𝗲 𝗵𝘂𝗺𝗮𝗻 𝘀𝘁𝘂𝗱𝗲𝗻𝘁𝘀 ...

𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗻𝗲𝗲𝗱 𝗹𝗮𝗽𝘁𝗼𝗽𝘀 - 𝗹𝗶𝗸𝗲 𝗵𝘂𝗺𝗮𝗻 𝘀𝘁𝘂𝗱𝗲𝗻𝘁𝘀 …

This week I had the pleasure of listening to Matt Bell’s presentation on Claude models. He covered the short but impressive history of Claude and the current state of affairs.

🎯 𝗧𝘄𝗼 𝗾𝘂𝗼𝘁𝗲𝘀 𝘀𝘁𝗼𝗼𝗱 𝗼𝘂𝘁 𝘁𝗼 𝗺𝗲:

 1️⃣ “𝗟𝗶𝗸𝗲 𝘆𝗼𝘂 𝗮𝗹𝘄𝗮𝘆𝘀 𝗴𝗶𝘃𝗲 𝘆𝗼𝘂𝗿 𝗲𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀 𝗮 𝗹𝗮𝗽𝘁𝗼𝗽, 𝘆𝗼𝘂 𝘀𝗵𝗼𝘂𝗹𝗱 𝗴𝗶𝘃𝗲 𝗖𝗹𝗮𝘂𝗱𝗲 𝗮 𝘃𝗶𝗿𝘁𝘂𝗮𝗹 𝗺𝗮𝗰𝗵𝗶𝗻𝗲.”

 2️⃣ “𝗪𝗲’𝗿𝗲 𝗱𝗼𝗶𝗻𝗴 𝘄𝗲𝗹𝗹 𝘄𝗶𝘁𝗵 𝗺𝗼𝗱𝗲𝗹 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲, 𝗯𝘂𝘁 𝘂𝘀𝗶𝗻𝗴 𝘁𝗼𝗼𝗹𝘀 𝗮𝗻𝗱 𝗼𝘁𝗵𝗲𝗿 𝗮𝗱𝗱-𝗼𝗻𝘀 𝗯𝗿𝗶𝗻𝗴𝘀 𝗺𝘂𝗰𝗵 𝗺𝗼𝗿𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗯𝗼𝗼𝘀𝘁.”

🇩🇪 𝗦𝗺𝗮𝗹𝗹 𝗚𝗲𝗿𝗺𝗮𝗻 𝗱𝗲𝘁𝗼𝘂𝗿: The first quote is fascinating. Here in Germany, we still struggle to equip students with laptops or tablets. There are ongoing debates about security, media impact, and the right age for device access. Yet I’m amazed watching my kids naturally complete schoolwork on their devices.

⚡ This isn’t just about competing with students from other countries anymore—agents are changing the game entirely.

🚀 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗯𝗿𝗲𝗮𝗸𝘁𝗵𝗿𝗼𝘂𝗴𝗵: The second quote reinforces the first. While model performance steadily improves, Anthropic sees far greater gains by providing agents with compute and coding capabilities. Security and scalability remain critical, making solutions like Amazon Bedrock AgentCore increasingly vital.

🧩 𝗠𝗼𝗱𝘂𝗹𝗮𝗿 𝗶𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻: This connects to Anthropic’s new Skills for Agents—modular folders containing knowledge and instructions that Claude loads on-demand for specific tasks. When prompted, Claude identifies relevant skills, loads necessary information, and executes tasks including API calls and code execution. This modular approach enables composability (multiple skills working together) and portability across Claude environments.

 I haven’t tested this yet, but it sounds like the right direction. It makes both building and executing agents more efficient. Instead of agents figuring out solutions from scratch or generating new code, they can reuse existing, modularized skills.

⚠️ This is crucial as we shift functionality from build time (developers create integrations) to runtime (agents solve problems dynamically)—currently a massive waste of tokens, compute, and energy.

Martin Karrer, Martin Sakowski - kind of what we discussed in your Agent Experience talk, about shifting efforts back to build time.

🌐 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱𝗶𝘇𝗮𝘁𝗶𝗼𝗻: For widespread adoption, the industry needs to agree on standards across model ecosystems. I’m glad that Maximilian Schellhorn has raised this in the Q&A, and the evolution of MCP shows it’s possible.

🔍 Next steps: I’m now exploring the Claude SDK’s compatibility with Amazon Bedrock. What’s your experience with this?