Claude Managed Agents: Build Production AI Agents in Days

TL;DR

On April 8, 2026, Anthropic launched Claude Managed Agents in public beta — a hosted platform that handles sandboxing, state management, tool execution, and error recovery so developers can ship autonomous AI agents in days instead of months. Pricing follows standard Claude API token rates plus $0.08 per active session-hour. Early adopters include Notion, Rakuten, Asana, Sentry, and Vibecode, with Allianz deploying custom agents across insurance operations. The architecture separates sessions, harnesses, and sandboxes into independent components — a design Anthropic says cut time-to-first-token by roughly 60% at P50.¹²

What You'll Learn

• What Claude Managed Agents is and the problem it solves
• How the session–harness–sandbox architecture works under the hood
• Pricing breakdown and what "active session-hour" means
• Which companies are already using it and for what
• How Managed Agents compares to OpenAI Agents SDK, Google ADK, and open-source frameworks
• What features are still in research preview and when to expect them

The Problem: Agent Infrastructure Is Expensive to Build

Building a production AI agent requires more than a good model. Developers need secure sandboxing, credential management, long-running session persistence, error recovery, tool orchestration, and tracing — infrastructure that typically takes months to build and maintain.¹

Industry analysis suggests that enterprises consistently underestimate the infrastructure investment required for production agent deployments — often by 40–60% relative to the agent development cost itself. Observability, identity management, and integration layers consume more resources than the agent implementation.³

Claude Managed Agents addresses this gap directly: it is a hosted service in the Claude Platform that runs long-horizon agents on your behalf through a small set of APIs.¹

How Managed Agents Works

Core Concepts

The platform is built around four primitives:⁴

Agent — the model (currently Claude models only), system prompt, tools, MCP servers, and skills that define what the agent can do.

Environment — a configured container template specifying pre-installed packages (Python, Node.js, Go, and others), network access rules, and mounted files.

Session — a running agent instance performing a specific task within an environment, generating outputs that persist even through disconnections.

Events — messages exchanged between your application and the agent, including user turns, tool results, and status updates. Events stream back via server-sent events (SSE).⁴

The Brain-and-Hands Architecture

Under the hood, Anthropic's engineering team decoupled the agent into three independent components inspired by operating system design — what they call "virtualizing the components of an agent."²

Session (the state) stores an append-only event log durably, independent of the harness or sandbox. If the harness crashes, no work is lost because the session persists outside it. Developers can query the event log with positional slices — picking up from the last read point, rewinding to see the lead-up to a decision, or re-reading context before taking action.²

Harness (the brain) is a stateless orchestration layer. It calls Claude, routes tool results, and manages context transformations. Because it is stateless, it can be replaced instantly: a crash is caught as a tool-call error, a new harness wakes with wake(sessionId), and execution resumes from the last event.²

Sandbox (the hands) is the execution environment where code runs, files are edited, and tools are invoked. Sandboxes are provisioned on demand via a provision({resources}) call and are treated as disposable — replaced rather than repaired. Each tool is exposed through a uniform execute(name, input) → string interface, whether it is a container, an MCP server, or a custom integration.²

This separation matters for performance. In Anthropic's original coupled design, the session, harness, and sandbox lived inside a single container. When it failed, the session was lost. After decoupling, P50 time-to-first-token dropped roughly 60%, and P95 dropped over 90% — because inference no longer waits for container startup.²

Security Boundaries

Credentials never reach the code-execution sandbox. Git tokens are injected during initialization but remain unavailable to generated code. OAuth tokens sit in a secure vault, accessed only through an MCP proxy that maps the session token to the right credentials. Claude never sees raw authentication material.²

Built-In Tools and MCP Support

Managed Agents gives Claude access to several built-in tool categories:⁴

Bash — shell commands run inside the container. File operations — read, write, edit, glob, and grep across the container's file system. Web search and fetch — search the web and retrieve content from URLs. MCP servers — connect to any external tool provider using the Model Context Protocol.

MCP integration is handled through a dedicated proxy. The proxy fetches authentication from the vault, calls the MCP server, and returns results — all without the agent or sandbox ever seeing raw credentials.²

This tool-use-first design means any custom tool, any MCP server, and any Anthropic-provided tool looks identical to the harness: a name and input go in, a string comes back.²

Pricing

Claude Managed Agents follows a consumption model:¹

Standard Claude Platform token rates apply (the same rates you pay for the Messages API). On top of that, there is a $0.08 per session-hour charge for active runtime. "Active" means the session is processing work — idle sessions do not accrue charges. When agents perform web searches through the platform, Anthropic charges $10 per 1,000 searches on top of model usage costs.⁵

For context: an agent running Claude Sonnet 4.6 for one hour of active work would cost roughly $0.08 in session fees plus whatever token costs the task generates. This compares to the alternative of provisioning, securing, and maintaining your own container infrastructure — which Anthropic argues costs orders of magnitude more when accounting for engineering time.¹

Who Is Already Using It

Several high-profile companies are building on Managed Agents:¹

Notion launched Custom Agents publicly in February 2026 and is now integrating Claude Managed Agents into the platform (currently in private alpha) for code shipping and content creation workflows. Rakuten deploys enterprise agents across product, sales, marketing, and finance teams via Slack and Microsoft Teams. Asana has integrated AI Teammates — collaborative agents that work within Asana projects. Sentry built a debugging agent that pairs with patch-writing and PR-opening capabilities. Vibecode reports that Managed Agents enables customers to deploy applications ten times faster.¹

Beyond these, Allianz — the global insurance conglomerate — signed a partnership with Anthropic in January 2026 that includes building custom AI agents for multistep workflows with a human in the loop, deploying Claude Code for all technical teams, and implementing an AI logging system for regulatory transparency.⁶

Research Preview Features

Three capabilities remain in research preview, requiring a separate access request:⁴

Multi-agent coordination — an orchestrator agent can spawn and coordinate multiple sub-agents in parallel, with Managed Agents handling communication and state sharing between them.

Advanced memory tooling — persistent memory that agents can read and write across sessions, enabling agents to build institutional knowledge over time.

Outcomes (self-evaluation) — developers define success criteria for a task, and Claude self-evaluates and iterates until it meets them, adding a goal-directed quality-control loop to the agent pipeline.

Anthropic has not announced a timeline for graduating these features to general availability.⁴

How It Compares to Alternatives

The agentic AI landscape in 2026 offers several paths to building production agents. Here is how they differ:⁷⁸

Claude Managed Agents is a fully hosted service. Developers define an agent, point it at an environment, and start sessions. The platform handles orchestration, sandboxing, and recovery. The trade-off is model lock-in: only Claude models run in the harness. If you later want to switch models, you rebuild the orchestration layer.⁷

OpenAI Agents SDK takes an open-source, code-first approach. Its core abstraction is the handoff — agents transfer control to each other explicitly, carrying conversation context through the transition. A working multi-agent system can be defined in under twenty lines of Python. The SDK supports non-OpenAI models through custom providers, though it is optimized for OpenAI's own models.⁸

Google Agent Development Kit (ADK) provides a hierarchical agent tree with deep integration into Google Cloud (Vertex AI Agent Engine) and Google's search infrastructure. Its differentiator is grounding — agents can search Google in real time and base responses on cited information, which addresses hallucination at the infrastructure level.⁸

LangGraph leads on production maturity among open-source frameworks, with built-in checkpointing, time-travel debugging, and LangSmith integration for observability. It supports any model but requires learning its graph abstraction — expect one to two weeks before a team is productive.⁷

CrewAI offers the lowest barrier to entry with a role-based DSL and processes over 12 million executions per day as of early 2026. It has raised $18 million in funding led by Insight Partners and reports 60% Fortune 500 adoption.⁷

The key distinction is hosted versus self-managed. Managed Agents eliminates infrastructure work but binds you to Claude. The SDKs and frameworks give you flexibility but require you to build and maintain the operational layer yourself.

What This Means for Developers

Claude Managed Agents represents Anthropic's bet that most teams building AI agents do not want to be in the infrastructure business. The platform absorbs the operational complexity — sandboxing, permissions, state management, error recovery — and exposes a clean API surface.

For teams already building on Claude Code or the Claude Agent SDK, Managed Agents provides a natural upgrade path to production deployment. For teams evaluating the MCP ecosystem, the platform's first-class MCP support means existing MCP servers plug in without modification.

The $0.08-per-session-hour pricing is aggressive enough that many teams will find it cheaper than maintaining their own agent infrastructure, especially when factoring in engineering time for security, recovery, and scaling.

The open question is whether model lock-in matters. In a landscape where the performance gap between frontier models narrows with each release, betting your agent infrastructure on a single provider is a real trade-off. Anthropic is clearly betting that Claude's capabilities — extended thinking, computer use, and deep MCP integration — are compelling enough to justify it.⁸

Getting Started

Claude Managed Agents is available today in public beta. To start building, you need a Claude API key and the managed-agents-2026-04-01 beta header on all requests (the SDK sets this automatically). The service is enabled by default for all API accounts — no waitlist required for the core platform.⁴

Multi-agent coordination, memory, and outcomes (self-evaluation) require a separate access request through Anthropic's form.⁴

Bottom Line

Claude Managed Agents is Anthropic's infrastructure play for the agentic AI era. By handling sandboxing, state management, and tool execution as a managed service, it removes the primary barrier that keeps most agent prototypes from reaching production. The brain-and-hands architecture is technically sound, the early customer list is strong, and the $0.08-per-session-hour pricing undercuts the true cost of self-hosted alternatives. The trade-off is model lock-in — a bet that Claude's capabilities justify the commitment. For teams already in the Anthropic ecosystem, this is a straightforward upgrade. For everyone else, it is a compelling argument to evaluate one.