Decouple Brain, Hands and Session
Separate the model loop (brain), the sandbox (hands) and the durable session log so each can fail and be replaced on its own; a dead container surfaces as a tool error and is re-provisioned, not nursed.
Evidence: moderate
Anthropic's first managed-agent design put the loop, the tools and the session together in one container, and that container became a "pet". A stuck session looked the same whether the cause was a harness bug, a dropped event or a dead container, and debugging meant opening a shell inside a container that held user data.
The redesign splits the agent into three virtualised parts. The session is an
append-only event log. The harness is the loop that calls the model and routes
tool calls. The sandbox is the execution environment. Once separated, the
container becomes cattle: it is called like any tool through an
execute(name, input) interface, its death reaches the model as an ordinary
tool error (see
failures returned as actionable feedback),
and a replacement is provisioned on demand. The harness is cattle too, because
recovery wakes a session by ID and resumes from the last event it wrote.
Credentials are kept out of reach of the sandbox entirely, which the separation
makes possible (see
credentials stay outside the sandbox).
Anthropic reports that the decoupling cut median time to first token by about
60% and 95th-percentile time by more than 90%.
This rests on one vendor's account of its own platform, and the latency figures are its own.
The pattern transfers to any containerised agent deployment: treat agent containers as disposable, keep the event log outside them in a session log outside the context window and a durable store, and let the loop crash and resume like any other replaceable part, in the spirit of journal-then-replay. There is a limit for agents that operate infrastructure: a worker cannot outlive the host or daemon failure it is meant to repair, which is one reason to prefer an external supervisor over self-diagnosis.
Sources
- Anthropic Engineering, 8 Apr 2026 — https://www.anthropic.com/engineering/managed-agents