The Product Shift

• Managed agents—offered by companies like Anthropic, Google, AWS, Microsoft, and LangChain—are defined by a three-layer stack: the model, the harness loop, and the runtime sandbox.1:31
• By owning the harness and runtime, vendors solve complex engineering tasks like credential handling, state persistence, and observability, effectively turning agents into long-lived async processes rather than simple request-response chats.2:06

Economics and Lock-In

• Managed agents are lucrative for vendors because they drive massive token consumption, with some interactions reaching up to 3 million tokens, often paired with session-hour fees.5:26
• Vendor lock-in occurs at different stack levels: model providers like Anthropic restrict you to their proprietary suites, cloud providers like AWS anchor you to their proprietary execution runtime, and frameworks like LangChain bind you to their observability control plane in LangSmith.6:40

Practical Constraints

• The primary barrier for enterprise adoption is data governance, as server-side persistence and cloud-based sandboxes often necessitate data leaving an organization's internal infrastructure.9:21
• If your product logic resides in the agent loop or requires zero-data-retention, these managed offerings are likely non-starters, necessitating a self-hosted or containerized approach using tools like Docker.