Scaling Managed Agents: Decoupling the brain from the hands

**The gist** Anthropic describes **Managed Agents**, a hosted service for long-horizon agent work that splits an agent into three virtualized parts: session log, harness, and sandbox. Harnesses run **stateless**, calling sandboxes through a generic **execute()** interface, so containers become disposable. Because the brain no longer waits on container provisioning, **p50 time-to-first-token dropped about 60%** and p95 over 90%.

**Why it matters** The pattern is reusable if you run your own agents: keep durable state in a session log outside the context window instead of irreversibly compacting it, keep credentials out of sandboxes (OAuth tokens stay in **external vaults behind MCP proxies**), and let **one brain drive multiple execution environments**. Harnesses can also deploy inside your own VPC.

**Watch out** This is architecture, not benchmark: no **cost comparisons**, scaling limits, or **session persistence guarantees** are given, and failure recovery is described in principle rather than measured.