Scaling Managed Agents: Decoupling the brain from the hands
02:00 · April 8, 2026 · Antropic Engineering Blog
Harnesses encode assumptions that go stale as models improve. Managed Agents—our hosted service for long-horizon agent work—is built around interfaces that stay stable as harnesses change.
Summary
Anthropic’s Managed Agents service provides a hosted platform for running long-horizon agents built around Claude. Rather than embedding the model, its control loop, and execution resources in a single container, the architecture separates three core components: the session, which stores an append-only log of events; the harness, which orchestrates calls to Claude and routes tool use; and the sandbox or other tools that perform actions. These components communicate through narrow, stable interfaces so that each can be replaced or scaled independently as models and infrastructure change.
The earlier coupled design placed the harness inside the same container as the sandbox. While this simplified file operations, it created operational problems. A failing container could lose the entire session, and debugging required direct access to an environment that might contain customer data. Network integration was equally rigid: connecting Claude to resources inside a customer VPC required either network peering or running the harness in the customer’s own environment. Both issues stemmed from assumptions about co-location that became liabilities as requirements grew.
Decoupling the harness from the sandbox treats execution environments as ordinary tools invoked through a simple execute interface. When a container fails, the harness records the error and can restart a fresh instance from a standard recipe. The session log lives outside the harness, allowing a new harness instance to resume from the last recorded event. Security improves because credentials never enter the sandbox; tokens are either injected at initialization or fetched through a proxy that keeps them out of Claude’s reach.
Context handling also benefits from the separation. The durable session log serves as an external store that the harness can query with positional slices, supporting compaction, trimming, or selective replay without irreversible loss of information. This design keeps context-management logic inside the harness, where it can evolve with future models, while the session guarantees only durability and accessibility.
The same interfaces enable practical scaling. Multiple stateless harness instances can connect to sandboxes only when needed, cutting time-to-first-token substantially. Individual brains can also coordinate several distinct execution environments, and teams can attach resources inside their own VPCs without changing the harness. By treating the session, harness, and sandbox as interchangeable abstractions, Managed Agents aims to accommodate harnesses and tools that have not yet been written.
Why it matters
Highly actionable for Product Teams and Builders implementing agent workflows with Claude, including code-level interface patterns, security mitigations, and performance gains like reduced TTFT. Directly addresses model updates, harness evolution, and production observability.
