Summary
Highlights
The Rise of Managed Agents00:00:00
Major tech companies including Anthropic, Google, AWS, Microsoft, and LangChain have recently launched managed agent products. These services go beyond simple model interaction by providing a managed harness and runtime sandbox.
Architecture of a Managed Agent00:01:28
A managed agent consists of three layers: the AI model, the harness (which handles loops, tools, and context management), and the runtime/sandbox (which holds credentials and provides observability). Providers typically own the harness and runtime layers.
Deployment Models and Concepts00:02:45
Managed agents operate in two main flavors: platforms where the provider manages the loop, and platforms where the user brings their own code to run in a provided sandbox. Concepts like sessions, environments, and events are central to these managed systems.
Economics and Vendor Lock-in00:05:25
Managed agents are high-token-usage products that shift costs toward session-based billing. These services often serve as a tool for vendor lock-in, requiring users to stay within a specific ecosystem of models, infrastructure, or observability platforms like LangSmith.
Strategic Considerations for Choosing00:08:41
To decide between managed systems and self-hosting, consider requirements for persistent state, sandbox security, data privacy, and zero-day retention. Enterprises with strict data residency requirements may find current managed agent offerings unsuitable.