Add note: Managed Agents architecture, unpacked

[herohua]

Summary

New tech note summarizing and analyzing Anthropic's Scaling Managed Agents, written for engineers with distributed-systems background and working knowledge of agent runtimes.

• Walks the article's redesign as one move (externalize state) applied recursively at three layers — sandbox, harness, credentials.
• Reads the resulting topology through an OS lens (process / WAL / syscall / capability / SCM_RIGHTS) to explain why the shape hangs together, not just what it is.
• Appendix unpacks three things the article folds together: harness-is-two-layers (per-step reducer vs. session driver), the credential delegation threat model, and where the MCP client most likely lives.
• Two mermaid diagrams: before (one container, one pet) and after (trust-zone subgraphs, session log off the request path).

Test plan

• Render check on github.io after merge — note: mermaid is not currently wired into the Jekyll theme, so the diagrams will display as raw code blocks until a mermaid.min.js include is added to the layout. Source view and GitHub UI render them natively.
• Front-matter fields match repo convention (title, date, tags, publish: true).
• No external assets referenced.

🤖 Generated with Claude Code