Why Anthropic’s Managed Agents Redefine AI Agent Runtime

Managed Agents Overview

Claude Managed Agents are defined by four core objects:

Agent : the model, system prompt, tool definitions, MCP servers and skills.

Environment : a container template that specifies installed packages, network access, and mounted files.

Session : a running instance of an agent that preserves history across multiple turns.

Events : the stream of interactions between your application and the agent (user messages, tool results, state updates).

Architecture Decoupling

Anthropic splits the runtime into three independent layers:

brain : Claude plus the orchestration harness. This component is stateless and can be restarted without losing session history.

hands : the sandbox and any attached tools. Treated as a generic tool interface.

session : a persistent event log stored outside the harness.

This "brain‑hands" separation removes the single‑point‑of‑failure that existed when the session, harness and sandbox were bundled in a single container.

Key Benefits

1. Fault isolation without container rescue

The sandbox implements a simple contract: execute(name, input) → string If the sandbox crashes, the harness records a tool‑call failure; Claude can retry or the system can provision a fresh environment.

2. Stateless harness with recoverable logs

Session logs are stored externally, so the orchestrator can be stateless. A crashed brain process can be revived with wake(sessionId) and continue processing events via getEvents().

3. Secure credential handling

Two patterns prevent tokens from ever reaching agent code:

Configure Git tokens during environment initialization so that git pull/push works, but the token is never exposed to the agent.

Store OAuth/MCP credentials in an external vault and access them through a proxy.

4. Session separate from context window

The context window acts as a temporary "workbench" for the current inference step, while the session log is a durable "repository" of all events. The workbench can be cleared at any time; the repository persists, allowing Claude to fetch needed information from the event stream instead of over‑filling the model's context.

Performance Gains

Anthropic reports that extracting the brain from the container reduces the 50th‑percentile time‑to‑first‑token (TTFT) by roughly 60 % and the 95th‑percentile by over 90 %. The brain can start immediately, and the hands are only provisioned when required.

Getting Started

The quick‑start flow consists of the following steps:

Create an Agent.

Create an Environment.

Create a Session.

Send a user.message event to the session.

Stream subsequent agent events.

Example CLI commands (the beta header anthropic-beta: managed-agents-2026-04-01 is required; SDKs add it automatically):

ant beta:agents create \
  --name "Coding Assistant" \
  --model claude-sonnet-4-6 \
  --system "You are a helpful coding assistant." \
  --tool '{type: agent_toolset_20260401}'

ant beta:environments create \
  --name "quickstart-env" \
  --config '{type: cloud, networking: {type: unrestricted}}'

Typical Use Cases

Long‑running coding, research, or operations tasks that need stateful continuity.

Scenarios requiring reliable recovery, retry, and full event history.

Integrations with custom MCP services, VPC‑bound resources, or external toolchains.

Projects where you prefer not to maintain a custom agent harness.

For short, single‑turn queries or simple tool calls, the standard Messages API remains a lighter option.