Claude Managed Agents

Claude Managed Agents的代理循环（Agent Loop）仍运行在Anthropic基础设施上，而工具执行迁移到企业自己配置的环境中

顾问策略（Advisor Strategy）是一种多Agent协作模式，允许主Agent在执行任务时咨询其他专业Agent的意见

Anthropic承诺所有新功能在原生Claude API上线的同一天也会同步部署到Managed Agents服务中

The code execution capability of Claude Managed Agents allows Agents to directly write and execute code to complete tasks such as data analysis, mathematical computation, and file processing.

Managed Agents' Agent model options include Opus, Sonnet, and Haiku, and the basic tool set typically includes bash, read, write, and grep.

Anthropic's initial Managed Agents architecture put the session log, Harness, and sandbox execution all in the same container, causing debugging difficulties, restricted data compliance, and rigid network topology.

The refactored Managed Agents is split into three interfaces: an append-only session log (Session), the Agent loop of Harness plus Claude (Brain), and sandboxes and tool calls provisioned on demand and destroyed after the task ends (Hands).

In the new architecture, the sandbox no longer holds credentials; requests to external resources go through a proxy, which injects the token before calling external endpoints or MCP, and the model never sees the credentials.

Because inference calls can start immediately without waiting for the container sandbox to be ready, the new architecture reduced first-token latency by about 90%.

The append-only event log allows component state to be rebuilt by replaying the log after a crash, and decouples the context compression strategy from persistent storage.

The Managed Agents API flow is: first use v1 environments to create an environment template and configure network policies and preinstalled packages, then create an Agent (specifying name, model, and system prompt and mounting a tool set), and finally create a Session with initial state idle using the Agent ID and environment ID.

A memory store shared across multiple sessions allows different sessions to write and read knowledge about financial instruments, avoiding duplicate research.

The Dreaming feature analyzes transcripts and memories across multiple sessions and automatically consolidates and updates the memory store, so that subsequent agents obtain more accurate information.

Vaults allow developers to define secrets; only placeholders are injected into the session, and the platform replaces them with real credentials at egress only when a pre-allowed domain is accessed.

The Deployments API supports running sessions on a schedule, suitable for long-term analyses that require periodic updates.

The hosted agent harness is inspired by Claude Code but rebuilt for general-purpose scenarios; it and the Agent SDK are different products: the latter is a downloadable binary, while the former is a fully managed service that includes hosting, sandbox management, scaling and event log storage.

Session communication is done through an event stream: the frontend sends user.message events, and the backend listens to the SSE stream and pushes the agent's responses back to the browser in real time.

In multi-agent collaboration, sub-agents do not inherit the full conversation history of the main session; they only copy the tool descriptions, schemas and prompts, to avoid context bloat and retain the capabilities needed for the task.

Skills are living artifacts that can be updated: after support for a new form type is added, new sessions that reference that skill automatically get the update.

The current platform does not provide session rollback or forking; users can only steer the agent through new events and cannot return to a historical point to retry.