Sources for Anthropic Managed Agents deep dive¶

Directly reviewed¶

1. https://www.anthropic.com/engineering/managed-agents
   - Title: Scaling Managed Agents: Decoupling the brain from the hands
   - Notes: Main engineering article. Core source for architecture, abstractions, performance, failure recovery, and security design.

2. https://platform.claude.com/docs/en/managed-agents/overview
   - Title: Claude Managed Agents overview
   - Notes: Product/docs layer. Used to map engineering abstractions to public product concepts: agent, environment, session, events.

3. https://www.anthropic.com/engineering/harness-design-long-running-apps
   - Title: Harness design for long-running application development
   - Notes: Used to understand the earlier “context anxiety” / context reset design that the Managed Agents article references as a harness assumption that later went stale.

Key extracted claims used in the write-up¶

• Harness assumptions can go stale as models improve.
• Old all-in-one container design turned the container into a fragile “pet”.
• Managed Agents decouples session, harness, and sandbox/tool execution.
• Session is durable event storage outside Claude’s context window.
• Sandboxes/tools can be treated through a stable execute(name, input) -> string style interface.
• Credentials should be isolated from generated code execution using remotes, vaults, and proxies.
• On-demand provisioning reduced time-to-first-token significantly.
• The system is designed for many brains and many hands rather than one model bound to one shell.