Claude Managed Agents: Cloud-Hosted Agent Platform at $0.08/hr

Anthropic Launches Claude Managed Agents

On April 8, 2026, Anthropic released Claude Managed Agents in public beta — a suite of composable APIs that let developers build and deploy cloud-hosted AI agents at scale. The platform abstracts away the operational complexity of running autonomous agents: container orchestration, state persistence, credential management, sandboxing, and error recovery are all handled by the platform, letting developers focus on the agent's behavior and user experience.

How It Works: Brain/Hands/Session Architecture

Managed Agents uses a three-part architecture:

• Brain — The Claude model handles reasoning, planning, and decision-making. Developers specify task descriptions, designate third-party tools (including MCP servers), and define security policies governing tool activation.
• Hands — Disposable Linux containers that execute tasks. Each container is ephemeral and isolated; if one crashes, a fresh container spins up automatically with no manual intervention.
• Session — A durable event log that sits outside both the brain and hands, recording every tool call, decision, and state change. This enables checkpoint-based crash recovery: if an agent fails mid-task, it can resume from the last checkpoint rather than starting over.

Sessions persist through disconnections, making it possible to run long-running autonomous workflows that survive network interruptions, container failures, and service restarts.

Key Capabilities

Production-Grade Sandboxing

Every agent runs in an isolated container with scoped permissions and identity management. Custom security rules govern which tools an agent can activate and under what conditions, giving enterprises fine-grained control over agent behavior.

Multi-Agent Coordination (Research Preview)

Agents can spawn sub-agents to break complex tasks into parallel subtasks. This multi-agent orchestration is currently available as a research preview — developers can request access through a form on the Claude Platform.

Self-Evaluation (Research Preview)

Claude can self-evaluate its work and iterate toward defined outcomes before returning results. In internal testing for structured file generation tasks, this approach improved outcome success by up to 10 percentage points over standard prompting, with the largest gains on difficult problems.

Tool Orchestration

The platform supports MCP servers and custom tools. Agents automatically determine which tools to use for specific prompts, and the orchestration layer handles tool discovery, invocation, and result handling.

Pricing

Managed Agents is billed at $0.08 per active session-hour plus standard Claude model token rates. Idle time does not count toward runtime charges — you only pay during active execution. An agent running continuously 24/7 costs approximately $58/month in runtime alone, before token costs.

Launch Partners

Five major companies launched integrations alongside the announcement:

• Notion — Custom Agents (private alpha) let teams delegate work to Claude directly from shared Notion task boards. Multiple agents can run in parallel, and outputs remain searchable and editable within the Notion workspace. Engineers can ship code and knowledge workers can produce websites and presentations without leaving Notion.
• Rakuten — Deployed enterprise agents across product, sales, marketing, and finance teams, accessible via Slack and Microsoft Teams. Went from prototype to production in under a week.
• Asana — AI Teammates that handle routine project management tasks collaboratively within Asana projects.
• Sentry — Pairs debugging agents with patch-writing capabilities, automating the path from bug detection to pull request generation.
• Vibecode — Powers AI-native app deployment, enabling 10x faster infrastructure setup for developer workflows.

Why It Matters

Building production-grade AI agents has required significant infrastructure investment: container orchestration, state management, security policies, observability, and crash recovery. Managed Agents packages all of this into a single API surface, dramatically lowering the barrier from prototype to production deployment. The $0.08/hr pricing makes it economically viable for high-volume agent workloads, and the multi-agent coordination preview signals Anthropic's direction toward complex, collaborative agent systems.