Anthropic launched Managed Agents on April 8 2026. Hosted infrastructure that handles sandboxing, state, permissions and error recovery for $0.08 per session hour. Notion, Rakuten, Sentry and Asana already in production.
Anthropic launched Claude Managed Agents in public beta on April 8 2026. The service absorbs four infrastructure layers that previously required developer teams weeks to build before writing a single line of business logic: secure sandboxing (isolated code execution), permissions management (granular access controls), state persistence (context maintained across long sessions), and error recovery (automatic retry with partial progress reporting). Internal benchmarks show task success rates improving by up to 10 percentage points over standard API prompting — the biggest gains on hard multi-step tasks where infrastructure failures previously caused silent drops. Notion, Rakuten, Sentry and Asana shipped production deployments on launch day. Access requires the managed-agents-2026-04-01 beta header, which the Claude SDK adds automatically.
Managed Agents offers two orchestration patterns. Agent Teams run multiple Claude instances with independent contexts, direct communication, and a shared task list — suited for complex parallel tasks. A Team might have one agent researching, one writing and one verifying simultaneously. Subagents run inside the main agent's session and only report results upward — cheaper in tokens, better for sequential tasks where each step depends on the previous. The $0.08 per session hour fee applies to both patterns on top of standard API token costs. Use Subagents as the default for most workflows. Reserve Agent Teams for tasks where genuine parallelism reduces total wall-clock time meaningfully.
The runtime fee compounds on long sessions. A 30-minute task costs $0.04 in runtime fees — negligible. An 8-hour research agent costs $0.64, which can exceed the token cost for shorter prompts. Enterprise users have flagged two concerns: runtime cost on long-running agents (4-8 hour sessions) where runtime exceeds token cost, and the absence of VPC peering — all traffic routes through Anthropic's public infrastructure. Compared to self-managed infrastructure: Managed Agents saves 1-2 weeks of initial engineering time plus ongoing maintenance. For teams running under 10,000 agent sessions per month, the managed service is typically cheaper when engineering cost is included. Above that volume, self-managed LangGraph or LangChain may be more economical.
Choose Claude Managed Agents when your stack is Claude-oriented, you want production-ready infrastructure in 30 minutes, and long asynchronous tasks are the primary use case. Choose LangGraph when you need model-agnostic flexibility, full orchestration control, or data cannot leave your infrastructure. Choose LangChain when you need the broadest integration ecosystem and are comfortable with higher engineering surface area. The practical observation: Managed Agents absorbs the 20% of LangChain and LangGraph that most production teams actually use. The remaining 80% — customisation, multi-model routing, on-premise deployment — keeps LangGraph relevant for sophisticated use cases.