Managed Agents - Don't Get Locked In

Tijdlijn van de aflevering

1. 0:00

   Managed agents have become a major new product category across nearly every major AI company in just five months.

   • Anthropic kicked off the managed agents trend in April, and most other companies have largely copied their ideas from the Anthropic offering.
   • Google launched anti-gravity agents in May followed by Gemini managed agents, AWS introduced agent core in June, Microsoft shipped foundry hosted agents in June, and LangChain released a managed deep agents public beta in mid-August.
   • OpenAI is expected to roll out its own version of managed agents at its September 29th dev day, completing the set of six major players in this space.
2. 0:50

   A managed agent consists of three layers: the model, the harness, and the runtime.

   • The bottom layer is the model itself, such as Opus and Claude for Anthropic or the latest GPT models on AWS and Azure.
   • The middle layer is the harness, which is the loop that calls models, picks and runs tools, feeds results back, and manages the context window.
   • The top layer is the runtime, the sandbox that holds credentials, does observability and tracing, and survives in the cloud by itself.
   • In managed agents, the provider owns levels two and three, and if you are Anthropic or OpenAI you also own the model layer.
3. 2:10

   Managed agents come in two flavors depending on whether the provider or the user brings the loop.

   • Flavor one means the provider owns the loop, so you send tasks and config without ever writing the loop or calling the model, as with Claude managed agents, Gemini agents, agent core harness, and LangChain managed deep agents.
   • Flavor two means you bring your own loop, perhaps written with LangGraph or ADK, and the provider hosts it in a container with per-session isolation while supplying runtime for tool execution and memory, as with foundry hosted agents and LangChain deployment.
   • Google and LangChain are hedging by offering both flavors, which makes sense for Google since DeepMind and Google Cloud build different things.

Kernbegrippen

• managed agents— The central topic — cloud-hosted agent services where the provider runs the harness and sandbox, not just the model.
• harness— The middle layer that runs the agent loop, calls models, picks and executes tools, and manages the context window.
• runtime and sandbox— The top layer where the loop actually runs, holding credentials, observability, and server-side state.

Opvallende citaten

• And I've got to say most the other companies have just copied their ideas from the anthropic offering.

  🔥

  — Reveals that the entire managed agents category is largely imitation rather than independent innovation, which is a surprising claim about a field with six major players.

  Ga naar 0:21

• a single anti-gravity agent interaction typically burns somewhere between 100k and 3 million tokens.

  🤯

  — The 30x range between 100k and 3 million tokens for a single interaction is genuinely shocking and reveals how uncontrolled server-side agent loops can be.

  Ga naar 5:39

Praktische conclusies