How Anthropic Built Managed Agents Infrastructure

Anthropic published the engineering architecture behind Managed Agents, separating session logs, inference harnesses, and execution sandboxes into independently scalable components. The design treats harnesses as stateless and containers as disposable - when something crashes, a new instance picks up from an append-only event log without losing progress. Credentials never enter sandboxes; OAuth tokens sit in external vaults and get injected through proxies. The performance payoff was significant: p50 time-to-first-token dropped 60 percent, p95 dropped over 90 percent, because inference no longer waits on container provisioning. The architecture borrows heavily from OS-level virtualization thinking, and the parallels to how Kubernetes decoupled compute from state feel intentional. Among AI labs shipping agent infrastructure, this is one of the more detailed public writeups on how the plumbing actually works.