Anthropic has introduced Managed Agents on its Claude platform, a managed execution layer designed to support the development and operation of agent-based workflows. The capability allows developers to define agent behavior, tools, and constraints while delegating runtime responsibilities such as orchestration, sandboxing, session state management, credential handling, and persistence to the platform.
The release targets production use cases where agents execute long-running, multi-step workflows involving external tools, error recovery, and continuity across sessions. Rather than requiring teams to build and maintain custom infrastructure for these concerns, Managed Agents expose APIs that standardize the deployment and execution of agent systems.
At a high level, Anthropic separates agent logic from execution infrastructure. Developers specify workflows and tool usage, while the platform provides a runtime responsible for secure execution, state handling, and operational guarantees. The company describes this as a meta-harness approach, where multiple agent workflows run on a shared execution substrate that handles common runtime concerns while preserving flexibility in agent design.
Managed Agents meta-harness architecture (Source: Anthropic Blog Post)
The system includes secure sandboxing for code execution, credential management for external systems, session continuity, and observability for debugging and auditing. Context management remains a consideration in long-running workflows, where systems determine what information to retain, summarize, or externalize beyond model context limits.
Radhika Menon, Senior Director AI at NTT DATA, described in linked in post:
All the infrastructure complexity that used to take months is now native to the platform. At 8 cents per session hour, you go from idea to production in days instead of months.
Other practitioners have raised concerns around ecosystem control and portability. Weilun Chen, Founder of Stealth, commented :
If the intention is to become a platform, the trajectory definition needs to be open source, and proposing a public open standard. But from what I read, this is a lock in into their SDK and their format.
Developers working with agent-based systems have commonly used external state stores and retrieval mechanisms to address limitations in model context windows. These approaches are used to maintain workflow continuity, support recovery from failures, and preserve traceability of agent behavior across long-running executions.
Context persistence and recovery are recurring considerations in stateful AI systems, particularly in workflows that extend beyond a single session. These requirements typically involve storing intermediate state outside the model context and retrieving it during execution to maintain consistency across steps.
Mufeez on X commented
Irreversible decisions to selectively retain or discard context can lead to failures
According to Anthropic, Managed Agents provide a meta-harness with interfaces for session state, sandboxed execution, and scaling across multiple agent workflows. The system separates runtime concerns such as state management and computation from agent logic, allowing general-purpose and task-specific agent configurations to execute on a shared runtime layer.