Summary
Sam Witteveen provides a comprehensive breakdown of the managed agents landscape, tracing how Anthropic, Google, AWS, Microsoft, LangChain, and the anticipated OpenAI offering (expected at Dev Day on September 29th) each shipped versions of the same fundamental product architecture within roughly five months of each other.
Witteveen explains managed agents as a three-layer stack — model, harness (the agentic loop), and runtime (the cloud sandbox) — and distinguishes two product flavors: provider-owns-the-loop offerings like Claude Managed Agents and Gemini Agents, versus bring-your-own-loop runtimes like Azure Foundry Hosted Agents and AWS Agent Core. He grounds the economics in real numbers: Google’s own documentation states that a single Antigravity agent interaction typically burns between 100,000 and 3 million tokens, making long-running agents the highest-token product a model company can sell — and the session-hour fee Anthropic charges on top of tokens signals where the whole category is heading.
The video closes on the lock-in implications: model companies like Anthropic and OpenAI use managed agent infrastructure to bind developers to their specific model families (Claude Fable, GPT), while cloud providers like AWS and Azure are betting that making the model interchangeable while owning the runtime is the stronger long-term play. For developers deciding where to build agent infrastructure today, Witteveen’s comparison provides a practical strategic map of the trade-offs between portability and convenience.
📺 Source: Sam Witteveen · Published September 13, 2026
🏷️ Format: Deep Dive
