Anthropic's Managed Agents Are Different (Here's Why)

하이라이트

• Claude Managed Agents allow users to build custom, secure, and production-ready agents using natural language without writing any code.

• The service operates on a session-based pricing model costing $0.08 per session hour plus standard Claude API token rates.

• A decoupled architecture separates the session memory (append-only logs) from the processing harness to allow for seamless recovery if a system fails.

• The Credential Vault secures sensitive data like API keys, ensuring that even the Claude model itself cannot access them until runtime.

• The managed environment supports integration with Slack, Notion, and various Model Context Protocol (MCP) servers for automated research and data retrieval.

타임라인

1. 00:00 - 01:27

   Introduction to Managed Infrastructure

   This service removes the need for users to manage their own Virtual Private Servers (VPS) or handle complex security protocols. While tools like OpenClaw offer high customization like Linux, Managed Agents function more like a closed ecosystem that prioritizes ease of use. Users define agent behavior through simple English prompts within the Claude console.

   • Managed agents represent the next evolution after the Claude Agent SDK.
   • The infrastructure handles security and scaling automatically to allow for production-ready deployment.
   • Natural language instructions replace traditional coding for agent creation and configuration.
2. 01:28 - 04:18

   Practical Implementation and Workflow

   In a test case involving medical data scraped from the NHS app, the agent successfully identifies allergies and provides health advice via Slack. The setup process involves three distinct steps: defining tools, organizing files/repos, and setting environment identities. Users can choose specific models like Claude 3.5 Sonnet to balance performance and cost during this configuration.

   • Agents can connect to private GitHub repositories to process sensitive medical or personal data.
   • The system suggests specific toolsets like read, glob, and grep while restricting unnecessary permissions like write or bash.
   • Configuration generates two primary files: a setup file for environment creation and a runtime file for communication.
3. 04:19 - 06:34

   Pricing and Session Management

   The pricing structure is bifurcated between token usage and active session time. While this adds a layer of cost beyond a monthly subscription, it allows for granular monitoring of how the agent interacts with tools like Slack and Bash. The UI allows for real-time versioning and adjustments to the system prompt or model selection without redeploying code.

   • Managed agents require Claude API billing rather than a standard Pro or Team subscription.
   • Idle sessions do not incur the $0.08 hourly charge.
   • Users can monitor live sessions and view full tool-call logs within the Claude console UI.
4. 06:35 - 11:00

   Technical Architecture and Security

   The orchestrator ensures high availability by spinning up new harnesses if one fails, using the append-only session logs to resume exactly where the previous instance stopped. Security is handled via a dedicated vault that keeps MCP and Anthropic API keys hidden from the model's logic. This design supports massive scaling and robust security for enterprise-level applications.

   • The system utilizes a three-part architecture consisting of a session, a harness, and an orchestrator.
   • The harness acts as a stateless router that executes tool calls and edits files in a sandbox.
   • Credential isolation prevents the model or the harness from knowing API keys until they are required at runtime.