Anthropic: Decoupling the Brain from the Hands Cut p95 Latency Over 90%

Anthropic's Scaling Managed Agents: Decoupling the Brain from the Hands (Lance Martin, Gabe Cemaj, Michael Cohen) is the architecture report, the first of the three reports. It splits a managed agent into three interfaces: a stateless brain (the harness loop that calls the model), replaceable hands (a sandbox called like any tool), and a durable session (an append-only log stored outside both). Provisioning the hands only when the agent needs them dropped p50 time-to-first-token roughly 60% and p95 over 90%.

What it establishes:

• Replaceable, not pets. A failed sandbox is re-provisioned; a failed harness reboots with wake(sessionId) and replays the session log, so a crash resumes instead of restarting.
• Credentials never enter the sandbox. They are bundled into a resource at init or held in a vault behind a proxy, so "the harness is never made aware of any credentials." Narrow token scoping is rejected on principle, since the model only gets better at misusing whatever it can reach.
• The session log is the audit trail. The same external, append-only log that lets a crashed agent resume is the evidence an auditor reads. The word "compliance" never appears in the post.
• What it leaves open. Attribution, log retention and access, data residency, and the compliance story a team owns.

Shopify reached the same brain/hands/durable-log split independently (Under the River), the strongest sign the architecture is convergent rather than Anthropic-specific. The guide builds on it in the reference architecture, the session, and credentials and secrets.