Scaling Managed Agents: Decoupling the Brain from the Hands
Source: Scaling Managed Agents — Decoupling the Brain from the Hands
Original: raw/docs/anthropic-managed-agents.md
URL: https://www.anthropic.com/engineering/managed-agents
Author: Anthropic Engineering
Summary
Anthropic's Managed Agents service redesigns agentic infrastructure around a single insight: harnesses encode assumptions that go stale as models improve. The fix is to decouple the three logical layers of an agent system so each can evolve or fail independently.
The three-layer model
| Layer | What it is |
|---|---|
| Brain | Claude + its harness (reasoning, tool-calling logic) |
| Hands | Sandboxes + tools that actually execute actions |
| Session | Append-only event log (persists outside both Brain and Hands) |
Key design decisions
Harness outside containers
Previously: harness ran inside the container alongside execution.
Now: harness lives outside, talks to containers via execute(name, input) → string.
Benefits:
- Container failures propagate as errors to Claude → Claude handles retries natively.
- No need to "nurse" failed containers back to health.
- Containers provisioned on-demand instead of pre-allocated per session → big latency win.
Durable session storage
The session log is stored outside the harness. On crash/restart:
- New harness calls
wake(sessionId). - Calls
getSession(id)to replay from last recorded event. - Full resume without state loss.
Security through separation
- Credentials never enter the sandbox where Claude-generated code runs.
- Git tokens: pre-wired at initialization.
- OAuth tokens (custom tools): stored in external vaults, accessed via proxies.
Performance results
| Metric | Improvement |
|---|---|
| p50 time-to-first-token | ~60% reduction |
| p95 time-to-first-token | >90% reduction |
Gains come primarily from on-demand container provisioning.
Scalability model
"Many brains, many hands": multiple Claude instances access multiple execution environments independently. No shared state between Brain instances → no single point of failure.
Related wiki pages
- Managed Agents Architecture — deep-dive concept page
- Harness Design Pattern — the harness-outside-container pattern
- Durable Session Storage — session log and recovery
- Anthropic — entity page