Claude Managed Agents? Everything You Need to Know

Building a production AI agent typically takes months, not because the AI is hard, but because the infrastructure around it is.

Claude Managed Agents, launched in public beta on April 8, 2026, removes that overhead entirely. Secure sandboxing, state management, credential handling, error recovery, and session persistence all come pre-built. You define what the agent does. Anthropic handles how it runs.

‍

Key Takeaways

• Infrastructure-as-a-service for AI agents: You define the agent's tasks, tools, and guardrails. Anthropic handles the secure execution environment, state management, and scaling.
• Solves the months-long deployment problem: Teams report going from prototype to production 10x faster, shipping in days where they previously projected weeks or months.
• Currently in public beta: Available to all Claude API account holders via the managed-agents-2026-04-01 beta header. Multi-agent coordination, self-evaluation, and memory are in separate research preview.
• Usage-based pricing: Standard API token costs apply, plus $0.08 per session-hour of active runtime. Idle time waiting for input does not count toward the charge.
• Five early adopters already live: Notion, Rakuten, Asana, Vibecode, and Sentry have all shipped production use cases, with Rakuten deploying each specialist agent within a week.
• Three paths exist: Managed Agents is the fully managed option. The Messages API and Claude Agent SDK are the self-managed alternatives, each with different trade-offs between control and overhead.

‍

What Is the Problem Claude Managed Agents Solves?

Building a production AI agent requires months of infrastructure work that has nothing to do with the agent's intelligence.

Secure sandboxed execution environments, state management across long-running sessions, credential handling, error recovery, context compaction, and end-to-end execution tracing all have to be built before a user sees anything.

Most teams write boilerplate they do not want to write: checking stop reasons, feeding tool results back into message arrays, managing context as it grows, deciding what to trim when limits approach.

• None of it is the actual problem: Every hour spent on infrastructure is an hour not spent on the agent behaviour that creates value for users.
• The cost gap was real: A $200 monthly Claude subscription was being used to run $1,000–$5,000 of agent compute before Anthropic closed this gap with a managed solution.
• The time gap was worse: Teams building legitimately on the raw API needed months before they could ship anything a real user would interact with.

Claude Managed Agents abstracts all of that infrastructure away. You define what the agent does. Anthropic handles how it runs.

‍

What Is Claude Managed Agents, Exactly?

Claude Managed Agents is a suite of composable APIs launched in public beta on April 8, 2026, for building and deploying cloud-hosted Claude agents at scale. It is available through the Claude Platform to all API account holders.

It provides a pre-built, configurable agent harness that runs in Anthropic's managed cloud infrastructure. Instead of building your own agent loop, tool execution layer, and runtime, you get a fully managed environment where Claude can read files, run commands, browse the web, and execute code securely.

Four core concepts from Anthropic's official documentation define how the service is structured:

• Agent: the model, system prompt, tools, MCP servers, and skills. Created once and referenced by ID across every session that uses it.
• Environment: a configured cloud container with pre-installed packages (Python, Node.js, Go), network access rules, and mounted files.
• Session: a running agent instance within an environment, performing a specific task and generating outputs.
• Events: messages exchanged between your application and the agent, including user turns, tool results, and status updates.

The harness includes built-in prompt caching, context compaction, and performance optimisations. Session tracing and integration analytics are built directly into the Claude Console, so you do not need to instrument your own logging to see what the agent is doing.

‍

How Does Claude Managed Agents Work?

Anthropic's documentation describes a five-step workflow that takes you from a definition to a running agent session. Each step is distinct and only needs to happen once except for starting sessions.

The workflow is intentionally linear so teams can get a first session running quickly and then build complexity from there.

• Step 1, Create an agent: Define the model, system prompt, available tools, MCP server connections, and any Agent Skills. This definition is stored and referenced by ID.
• Step 2, Create an environment: Configure the cloud container, including which language packages are pre-installed, network access rules, and any files to mount at session start.
• Step 3, Start a session: Launch a session referencing your agent and environment. The session persists through disconnections; progress and outputs are retained server-side automatically.
• Step 4, Send events and stream responses: Send user messages as events. Claude autonomously executes tools and streams results back via server-sent events (SSE). Full event history is persisted and can be fetched at any time.
• Step 5, Steer or interrupt: Send additional user events to guide the agent mid-execution, or interrupt it entirely to change direction without losing session state.

The orchestration harness handles tool calling decisions, context management, and error recovery automatically. You do not write or manage the agent loop.

A self-evaluation capability in research preview lets you define success criteria, with Claude iterating until it meets them, useful for tasks where "good enough" requires judgment rather than a binary pass or fail.

‍

What Tools Does the Agent Have Access To?

A managed agent is not limited to generating text. It can take actions in the real world through a set of built-in tools and extensible connections.

Every session has access to the same core tool set out of the box, with no additional configuration required.

• Bash: run shell commands in the managed container, including scripts, build processes, and data transformations.
• File operations: read, write, edit, glob, and grep files within the session container, covering the full lifecycle of file-based tasks.
• Web search and fetch: search the web and retrieve content from specific URLs, giving the agent access to live information during a session.
• MCP servers: connect to any external tool provider through the Model Context Protocol, the same protocol that powers Claude Code's external integrations.

MCP connectivity is where the extensibility becomes significant. Agents can connect to GitHub, Jira, Slack, Google Drive, or any custom tool your team has exposed as an MCP server.

Agent Skills, which are pre-built capability packages for PowerPoint, Excel, Word, and PDF file generation, can also be loaded dynamically when the agent needs to produce structured document outputs.

Multi-agent coordination, currently in research preview, allows agents to spin up and direct other agents to parallelise complex work. This is the capability behind Rakuten's parallel specialist agent deployments and Notion's parallel task execution workflows.

‍

How Does Claude Managed Agents Differ From the Messages API and the Agent SDK?

Anthropic now offers three distinct paths for building with Claude. Understanding which path fits your situation is more important than understanding any single path in detail.

The three paths exist because different teams have different constraints. Some need maximum control. Some need to move fast. Most fall somewhere in between.

• Messages API: direct model access. You build and manage your own agent loop, tool execution, state management, and infrastructure. Maximum control, maximum overhead. Best when you need custom agent loops or fine-grained architectural control that Anthropic's managed environment cannot provide.
• Claude Agent SDK: the infrastructure that powers Claude Code, exposed as a Python and TypeScript library. It manages the agent loop, context compaction, and built-in tool execution for you, but you still host and run it yourself. Best when you need programmable control without building the loop from scratch.
• Claude Managed Agents: fully managed cloud infrastructure. You define the agent. Anthropic runs it. No servers, no loop management, no sandboxing to build. Best for long-running tasks, asynchronous work, and production deployments where operational overhead is the constraint.

The practical distinction is this: if you need fine-grained architectural control, custom memory systems, or unusual tool patterns, the Messages API or Agent SDK give you that.

If you need to get to production without building infrastructure, Managed Agents is the right path.

‍

What Are Teams Already Building With It?

Five organisations were live in production at launch. Each one illustrates a different use case and a different kind of bottleneck that Managed Agents removed.

The range across Notion, Rakuten, Asana, Vibecode, and Sentry shows the breadth of what "long-running, tool-using agent" looks like across industries and product types.

• Notion: engineers and knowledge workers delegate open-ended tasks directly inside Notion workspaces through Custom Agents.
• Rakuten: deployed specialist agents across product, sales, marketing, finance, and HR. Agents plug into Slack and Teams, accept task assignments, and return completed deliverables including spreadsheets, slide decks, and apps in sandboxed environments. Each specialist agent was deployed within a week.
• Asana: built AI Teammates, collaborative agents that work alongside humans inside Asana projects, picking up tasks and drafting deliverables. CTO Amritansh Raghav confirmed Managed Agents accelerated development and freed the team to focus on enterprise-grade user experience rather than infrastructure.
• Vibecode: uses Managed Agents as the default integration for its prompt-to-deployed-app product. Before Managed Agents, users managed LLM sandboxes manually, a setup that took weeks to months. With Managed Agents it takes a few lines of code and deploys at least 10x faster.
• Sentry: paired their existing Seer debugging agent with a Claude-powered agent that writes the patch and opens the PR. A flagged bug now flows directly to a reviewable fix. The integration shipped in weeks rather than the months originally projected.

‍

What Does Claude Managed Agents Cost?

Pricing has two components: token costs and runtime costs. Both are usage-based with no platform fee or minimum commitment.

Token costs follow standard Anthropic API pricing. Sonnet 4.6 runs at $3 per million input tokens and $15 per million output tokens.

Opus 4.6 runs at $5 per million input tokens and $25 per million output tokens. Prompt caching applies at standard multipliers.

• Runtime charge: $0.08 per session-hour of active runtime, measured in milliseconds. This is the infrastructure cost for running the managed container during execution.
• What active runtime means: only time when the agent is actively executing, running code, calling tools, or processing responses. Idle time when the agent is waiting for user input or paused between tool calls does not count toward the runtime charge.
• Web search cost: $10 per 1,000 searches, consistent with standard API pricing. Input and output tokens consumed to process search results are billed separately at standard rates.
• Rate limits: 60 create requests per minute for agents, sessions, and environments. 600 read requests per minute for retrieve, list, and stream operations.

Access is enabled by default for all Claude API accounts. The beta header managed-agents-2026-04-01 is required on all requests.

Research preview features, including multi-agent coordination, outcome-based self-evaluation, and memory, require a separate access request at claude.com/form/claude-managed-agents.

‍

What Is Still in Research Preview and What That Means for Production?

Three features are in research preview at launch. They require a separate access request and are not available by default even to accounts with Managed Agents access.

Being clear about this matters. Teams who build expecting these features and discover they require separate access lose time they cannot recover.

• Outcomes and self-evaluation: define success criteria and Claude iterates until it meets them. Useful for tasks where passing or failing requires judgment rather than a binary check. In research preview.
• Multi-agent coordination: agents spinning up and directing other agents for parallelised complex work. The capability behind Rakuten's parallel deployments and Notion's parallel task execution. In research preview.
• Memory: persistent memory that survives across sessions. Currently in research preview, meaning cross-session memory requires custom implementation for most users building today.

The rest of the service, including sessions, built-in tools, MCP connections, Agent Skills, and standard orchestration, is in public beta and is considered stable enough for production use. Rate limits and pricing are established but may be refined as the beta period continues.

Teams can build and ship production agents today without research preview access. Multi-agent and outcome-based workflows require it, and access must be explicitly requested.

‍

Who Is Claude Managed Agents For?

Claude Managed Agents is designed for a specific kind of team with a specific kind of constraint. Matching that constraint is more important than any feature comparison.

If the bottleneck in your agent deployment is infrastructure time and overhead, Managed Agents directly solves it. If the bottleneck is architectural flexibility or proprietary orchestration requirements, the Messages API or Agent SDK is the better fit.

• Teams who want to ship agents fast: the primary audience. If the constraint is getting to production in weeks rather than months, Managed Agents removes the infrastructure bottleneck entirely.
• Companies embedding agents inside existing products: Notion, Asana, Sentry, and Vibecode are all embedding Claude-powered agents inside their own products. Managed Agents handles the infrastructure layer so their engineering teams can focus on the product experience.
• Enterprises running long-duration, asynchronous workflows: any workflow that runs for minutes or hours, involves multiple tool calls, and needs to persist through disconnections is a strong fit. Document processing, code analysis, multi-step data pipelines, and cross-functional automation all sit in this category.
• Teams not ready for a full infrastructure investment: for startups and smaller engineering teams, months of infrastructure work to build a production-grade agent loop is a genuine barrier. Managed Agents reduces that to a few lines of code.

Teams who should use the Messages API or Agent SDK instead are those who need custom orchestration logic, unusual memory architectures, deeply proprietary tool integrations, or fine-grained control over every step of the agent loop. The managed path is faster. The self-managed path is more flexible.

‍

Conclusion

Claude Managed Agents is Anthropic's answer to a specific and persistent problem: the gap between a working agent prototype and a production-grade deployment.

It does not make the agent smarter. It removes the months of infrastructure work that have kept most teams from shipping agents at all. For teams where that infrastructure gap is the real constraint, Managed Agents directly solves it. For teams who need architectural control a managed environment cannot provide, the Messages API and Agent SDK remain the right paths.

Access is available to all Claude API accounts today. The fastest path to evaluation is the Managed Agents quickstart at platform.claude.com/docs/en/managed-agents/quickstart, which takes minutes to run a first session rather than building infrastructure to find out whether the service fits.

‍

Building Agents on Claude Managed Agents? We Can Help You Ship Faster.

Most teams do not struggle with the idea of what they want to build. They struggle with getting from a working prototype to something that runs reliably in production, handles edge cases, and integrates cleanly with the rest of their stack.

At LowCode Agency, we are a strategic product team, not a dev shop. We help teams scope, design, and deploy production Claude agents using Managed Agents infrastructure, from defining agent architecture to MCP integration to testing and handoff.

• Agent architecture scoping: We map your workflow, define the agent's scope and tools, and design the session structure before any code is written.
• MCP integration: We connect your managed agent to your existing tools (GitHub, Jira, Slack, Google Drive, or custom internal systems) using the Model Context Protocol.
• Knowledge base and system prompt design: We write the system prompts and structure the knowledge inputs that determine how reliably your agent performs on real tasks.
• Production deployment: We configure environments, set up monitoring through Claude Console, and build the event handling layer between your application and the managed session.
• Multi-agent coordination design: For teams with research preview access, we design the specialist agent architecture and inter-agent task routing that complex parallel workflows require.
• Testing and calibration: We run your agent against real inputs, document failure modes, and refine prompts and tool logic before any user sees the output.
• Full product team: Strategy, design, development, and QA from a single team invested in your agent performing in production, not just getting deployed.

We have built 350+ products for clients including Coca-Cola, American Express, and Medtronic. We are an Anthropic Claude Partner Network member and have been working with Claude's API since early access.

If you are building on Claude Managed Agents and want to move from prototype to production without the trial-and-error cycle, let's scope it together.