Managed Agents Operations
Summary — Key Points of This Lesson
- Managed Agents is an architectural pattern for managing long-running, multi-step agent tasks. The core design separates the system into three layers: session (brain) · harness (brain) · sandbox (hands).
- Separating the harness from the container significantly reduces Time to First Token (TTFT) (source: Anthropic Engineering: Managed Agents).
- A recommended security design keeps credentials (API keys, etc.) separate from the sandbox, using them only during initialization.
- Fault-tolerant design — where a failed harness can be recovered from a new session ID — is a prerequisite for large-scale operations.
- The key to long-term operation is designing each component to be independently replaceable, accounting for the fact that model improvements will eventually make today's assumptions obsolete.
目次 (6)
What Are Managed Agents?
Managed Agents is an architectural design pattern for stably operating long-running, multi-step autonomous agent tasks at scale. The Anthropic engineering blog (Anthropic Engineering: Building effective managed agents) provides a detailed explanation.
For an in-depth architecture walkthrough, see also the Managed Agents Large-Scale Operations Guide.
Brain and Hands Separation — 3-Layer Architecture
The core of Managed Agents is separating the agent system into three independent components.
| Layer | Role | Analogy |
|---|---|---|
| Session | Append-only log of all events that have occurred | Long-term memory of the brain |
| Harness | Calls Claude and routes tool invocations | Prefrontal cortex of the brain (decision-making & direction) |
| Sandbox | Actual processing such as code execution and file operations | Hands (execution environment) |
This separation allows each layer to be replaced independently without affecting the others. For example, if the model improves and the harness assumptions become outdated, you can update only the harness while keeping the session and sandbox as-is.
Scalability and Fault-Tolerant Design
For large-scale Managed Agents, a configuration where multiple harnesses (brains) connect to multiple sandboxes (hands) is recommended. The main benefits are as follows:
- Fault tolerance — If a harness fails, a new harness can be launched from the session ID to continue.
- Parallel scaling — Multiple harnesses can execute tasks in parallel across independent sandboxes.
- TTFT improvement — Separating the harness from the container reduces wait time for the first token generation (Anthropic internal experiments showed approximately 60% reduction at p50 and over 90% at p95).
Security: Credential Separation
To operate Managed Agents safely, a design where credentials (API keys, authentication tokens, etc.) are not permanently resident inside the sandbox but are used only during the initialization phase is recommended. Configuring custom tools to access the credentials vault via MCP (Model Context Protocol) minimizes credential exposure to the sandbox.
The Design Philosophy of "Assumptions Becoming Obsolete"
An important premise in Managed Agents design is that "today's assumptions will become obsolete tomorrow as models improve."
For example, designing a harness based on the assumption that "Claude Sonnet 4.5 terminates tasks early when it senses context limitations" will break down once the model improves. Keeping components loosely coupled and designing them to be individually replaceable is the key to long-term stable operations.
The knowledge from Level 4 "Harness Design" and "Leveraging Sub-agents" forms the implementation foundation for Managed Agents.
Level 6 — And All 30 Lessons Complete
Congratulations. You have completed all 5 lessons of Level 6 "Business Use." Claude Cowork introduction · Excel integration · Team/Enterprise plans · ZDR · Managed Agents — you now have a comprehensive understanding of everything needed to integrate Claude into your business workflows.
And this is the final lesson of the 30-lesson Claude Master series. From Level 1 "Introduction" to Level 6 "Business Use," you have systematically learned everything from the fundamentals of Claude to its most advanced business applications. Going forward, deepen the knowledge you have gained through hands-on practice in your actual work and projects.
