The agent, the workspace and the record should fail independently.
Anthropic detailed the architecture behind Managed Agents, separating the harness, execution sandbox and durable session history. The design addresses recovery, changing model capabilities and the boundary around credentials and actions.
Anthropic's April 8 engineering post describes how Managed Agents separates the model and harness from execution environments and a durable session log. The company explains that combining all three in a single container created problems when a container failed or customers needed a different infrastructure boundary.
With separate components, the workspace can be replaced and the harness can resume using stored events. The post also describes keeping credentials outside the sandbox that executes generated code. This is an architecture account from the provider, not an independent reliability benchmark.
Our view: ask recovery questions before autonomy questions. What happens if a process crashes halfway through a customer request? Can the system distinguish a failed action from a completed action whose response was lost? Is there enough evidence for a human to reconstruct what happened? A credible production agent needs clear answers before it receives consequential responsibilities.
recovery is a core capability of an agent, not an afterthought.
Read the primary source. Editorial interpretation is our own.