Claude Managed Agents

Learning Objectives

• Understand what Managed Agents offers and how it differs from Claude Code and Claude Cowork
• Identify the key capabilities of cloud-hosted agent deployment
• Evaluate when to use Managed Agents versus building agents with the Claude Agent SDK

What Are Claude Managed Agents?

Claude Managed Agents is a public beta product from Anthropic that provides composable APIs for building and deploying cloud-hosted AI agents at scale. Where Claude Code runs locally in a terminal and Cowork's agentic work runs on your desktop inside the Claude app, Managed Agents lets organizations deploy persistent, cloud-based AI agents through Anthropic's infrastructure.

The key distinction: with Claude Code or the Agent SDK, you build and host agents yourself. With Managed Agents, Anthropic handles the orchestration, scaling, monitoring, and infrastructure — you define what the agent does, and Anthropic runs it in the cloud.

How It Fits in Anthropic's Agent Stack

Core Capabilities

Composable Agent APIs

Managed Agents provides APIs for defining agents as composable building blocks:

• Define agent behavior — specify tools, instructions, and constraints through the API
• Chain agents together — compose multi-agent workflows where agents hand off to each other
• Persistent state — agents maintain state across sessions without you managing a database
• Automatic scaling — Anthropic handles scaling as demand grows

Cloud-Hosted Execution

Unlike self-hosted agents, Managed Agents runs on Anthropic's infrastructure:

• No server management — no need to provision, scale, or monitor compute resources
• Built-in monitoring — observe agent behavior, track tool usage, and audit decisions
• Reliability — Anthropic manages uptime, failover, and infrastructure maintenance
• Security — enterprise-grade security and data handling through Anthropic's platform

Enterprise Features

• Role-based access — control who can create, modify, and deploy agents
• Audit logging — full trail of agent actions for compliance
• Rate limiting and cost controls — set usage boundaries per agent or team
• Integration hooks — connect agents to internal systems via MCP or custom tools

Credentials Vault and NVIDIA OpenShell

Managed Agents runs the agent loop on a separate server from the sandbox where the work happens, and keeps passwords and access keys in a vault the agent never sees. On September 28, 2026 Anthropic announced that customers can pair it with NVIDIA OpenShell, open-source runtime software that blocks every action unless a rule allows it, checks each tool an agent tries to use, and logs every decision outside the agent itself. Used together, a company can limit what an agent can do, review what it did, and confirm the limits were in place.

The pairing arrived the same day NVIDIA launched its Open Agent Safety Platform, which adds a hardware monitor called Sentry running on a separate BlueField-4 data processing unit. Anthropic is among the platform's listed supporters. The layers are designed to work independently, so protection does not depend on the model behaving well, but they are new: no independent test of how well they contain a determined agent has been published.

Use Cases

Strengths

• Zero infrastructure management — Anthropic handles hosting, scaling, and monitoring
• Composable architecture — build complex workflows by chaining simple agent building blocks
• Persistent state — agents maintain context across sessions automatically
• Enterprise-ready — audit logging, access controls, and cost management from the start
• Powered by Claude — runs on the current Claude Fable, Opus, Sonnet and Haiku models

Limitations and Considerations

• Public beta — the product is still maturing; APIs and features may change
• Cloud-only — agents run on Anthropic's infrastructure, not on-premise (important for data sovereignty requirements)
• Pricing not finalized — beta pricing may differ from GA pricing
• Vendor dependency — deeper coupling with Anthropic's platform compared to self-hosted Agent SDK solutions
• Limited customization — less control over execution environment compared to self-hosting

Company Details

Related Tools

• Claude Opus 5.5 — Anthropic's current Opus-line model powering these agents, leading since September 22, 2026
• Claude Agent SDK — Self-hosted agent framework for custom deployments
• Claude Cowork — Desktop agent for knowledge work
• Claude Mythos 5.1 — Anthropic's most powerful model (invitation-only)

Key Takeaways

• Claude Managed Agents is Anthropic's public beta for deploying cloud-hosted AI agents at scale — you define the agent, Anthropic handles infrastructure, scaling, and monitoring
• Composable APIs allow chaining agents into multi-step workflows with persistent state across sessions
• Fills the gap between desktop agents (Claude Code/Cowork) and self-hosted agents (Agent SDK) — ideal for organizations that want agent capabilities without managing infrastructure
• Enterprise features include audit logging, role-based access, and cost controls
• Credentials sit in a vault the agent never sees, and since September 28, 2026 customers can add NVIDIA OpenShell rules that limit and log what an agent can reach
• Currently in public beta — APIs and pricing may evolve before general availability