Illustration by swanbase, based on claude.com/blog/claude-managed-agents and Anthropic's documentation
Claude Managed Agents, launched by Anthropic on April 8, 2026, is cloud infrastructure for deploying AI agents in production without having to handle sandboxing, state management, permissions, or long-running sessions yourself. You define what your agent should do, and Anthropic runs it. According to teams already in production quoted in the announcement, the result is a shift from months of infrastructure work to a few days of development. It's available in public beta on the Claude Platform. Pricing: standard Claude token rates plus $0.08 per active session-hour.
If you have a working prototype and production infrastructure is your next roadblock, Managed Agents lets you skip that whole cycle. If you need maximum control over your agent loop, the Messages API is still the better fit.
The infrastructure cycle: from months to days
The infrastructure you used to have to build
Putting an AI agent into production meant running a full stack yourself: a sandboxed environment for code execution, state management across sessions, a scoped permissions system, secure credentialing, checkpointing for long tasks, and end-to-end tracing to debug failures. Each layer is a project in its own right. And every model update means reworking your agent loops.
The upshot: weeks, sometimes months, before anything reaches production. That's engineering time that doesn't go into the product.
The stack Anthropic takes off your plate
All of the above. You define the tasks, the available tools, and the guardrails. Anthropic runs everything on its infrastructure. John Han, co-founder of a meeting-prep product, sums it up in the announcement: "3x faster to build a production-ready agent. From idea to production in a matter of days. We could focus on the product instead of the infrastructure."
Messages API or Claude Managed Agents: which one to choose
| Messages API | Claude Managed Agents | |
|---|---|---|
| Definition | Direct access to the model | Preconfigured agent harness, managed infrastructure |
| Best for | Custom agent loops, fine-grained control | Long-running tasks, asynchronous work |
| Infrastructure | You build and maintain it | Managed by Anthropic |
| Sessions | Stateless by default | Persistent, long-running (several hours) |
| Multi-agent | Implement it yourself | Built in (research preview) |
| Tracing | Local logs | Integrated in Claude Console |
| Model updates | Rework the agent loop | Handled by the harness |
Rule of thumb: if you're building a demo or an internal tool with a simple agent, the Messages API is enough and costs less. If you're shipping to users, with tasks that run for hours and need to survive disconnections, Managed Agents is the right infrastructure.
The 4 core concepts
Managed Agents is built on four concepts, each covered in a section below: agent, environment, session, and events.
Agent
The configuration: the Claude model, the system prompt, the tools, the MCP servers, and the skills you enable. You create an agent once and reference it by ID across every session.
Environment
A preconfigured cloud container: Python, Node.js, and Go packages pre-installed, network access rules, mounted files. You define the environment once, and every session starts in that clean, secure container.
Session
A running instance of your agent inside an environment. It can last a few minutes or several hours. State persists for the whole duration: the agent creates files, edits them, and reads them back in later steps. If your connection drops, the session keeps running on Anthropic's side and the outputs are waiting for you when you reconnect.
Events
A real-time trace of everything the agent does: every tool call, every decision, every intermediate output. You debug precisely, directly in Claude Console, without having to piece the history back together.
When an agent needs to route a task or check a tool call before it executes, Jev from TypeSafe AI adds a typed decision to the loop while leaving generation to Claude.
The layers Anthropic handles for you
Illustration by swanbase, based on claude.com/blog/claude-managed-agents, Anthropic
Sandboxing and scoped permissions
Code runs in an isolated container. The agent can't touch your machine, your file systems, or your credentials unless you explicitly configure it to. The permissions system has two modes: workspace (read-write inside the container) and internet (controlled network access). You set these permissions when you create the agent, with no sandboxing code to write or maintain.
Long-running sessions and disconnections
Agents run autonomously for hours. If your connection drops, the session carries on and the outputs persist. You pick up wherever the agent is, with no restart. A standard API call, by contrast, fails the moment the connection goes down.
Context engineering: how Anthropic solved "context anxiety"
Anthropic documents this technical detail on its engineering blog. Claude Sonnet 4.5 tended to wrap up tasks prematurely when it sensed it was nearing its context limit, a behavior dubbed "context anxiety". The agent would rush to a conclusion even when the task wasn't finished.
The fix: context resets in the harness. The harness manages the context window automatically, compresses the history when needed, and resumes cleanly. You don't have to deal with this in your own code, because the harness takes care of it.
In internal testing on structured file generation tasks, Anthropic measured a success rate up to 10 points higher than a standard prompting loop, with the biggest gains on the hardest tasks.
Multi-agent coordination
Child agents directed by a parent agent
Agents can spin up and direct other agents to parallelize complex work. A parent agent breaks a task down, delegates subtasks to child agents, and aggregates the results. Internally at Notion, dozens of tasks run in parallel while the team collaborates on the outputs.
How to request access
This feature is in research preview and available on request. The request form is linked from the Managed Agents announcement page on claude.com. Everything else in Managed Agents (single sessions, tooling, sandboxing) is available right now to every developer with API access.
Pricing in practice
Structure: tokens + $0.08 per session-hour
Two components:
- Claude tokens: standard Claude Platform rates, the same as regular API usage
- Session-hours: $0.08 per hour of active session
Real cost examples by use case
A code review agent that runs for 20 minutes: $0.027 in session-hours, plus the tokens consumed (which vary with the volume of code). A document-processing agent that runs for 3 hours: $0.24 in session-hours, plus tokens.
In most cases, tokens remain the dominant cost, and session-hours barely register for a well-designed agent. Weigh that $0.24 for three hours of agent runtime against the cost of an engineer spending a week setting up sandboxing, state management, and tracing.
One important point: pricing only counts active hours. An idle agent doesn't rack up session-hours.
Use cases already in production
Illustration by swanbase, based on claude.com/blog/claude-managed-agents, Anthropic
Notion, Rakuten, Asana, Vibecode, Sentry
These teams have been in production since the beta launched, according to Anthropic's announcement:
Notion lets users delegate work to Claude straight from the workspace: engineers hand off code, while other teams hand off slides and websites. Dozens of tasks run in parallel while the team collaborates on the results.
Rakuten deployed specialized agents for each function (product, sales, marketing, finance), connected to Slack and Teams. Employees assign them tasks and get deliverables back. Each agent took less than a week to deploy.
Asana has AI Teammates working alongside humans in Asana projects, picking up tasks and drafting deliverables. The team built its advanced features "significantly faster" than with a traditional setup.
Sentry chains its Seer agent, which detects a bug, with a Claude agent that writes the patch and opens the PR: from detected bug to review-ready fix in a single flow. The team shipped the integration in a few weeks.
Vibecode uses Managed Agents as its default integration to go from a prompt to a deployed app. Its users stand up this infrastructure at least 10 times faster than before, according to its co-founder quoted in the announcement.
How to get started in under 10 minutes
Via Claude Console
The fastest route is platform.claude.com/workspaces/default/agent-quickstart: a visual interface to configure an agent, test a session, and copy the agent ID into your code, with no command line needed for a first test.
Via CLI
Anthropic launched an ant CLI alongside Managed Agents:
brew install anthropic-cli
ant --version
Then create your first agent through the API:
curl https://api.anthropic.com/v1/agents \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-beta: managed-agents-2026-04-01" \
-d '{
"name": "code-reviewer",
"model": "claude-sonnet-4-6",
"system_prompt": "You are a code reviewer...",
"tools": [{"type": "agent_toolset_20260401"}]
}'
Via Claude Code
If you already use Claude Code, this is the fastest route of all. Type this directly into Claude Code:
start onboarding for managed agents in Claude API
The built-in skill walks you through the entire setup.
What kind of project is it for?
Managed Agents makes sense if:
- You're shipping an agent to your end users rather than a one-off internal tool
- Your tasks run longer than a few minutes and need to be reliable
- You have a small dev team and don't want a parallel infrastructure project
- You need multi-agent capabilities (research preview, on request)
The Messages API is still the better choice if:
- You're prototyping or running quick tests
- You need very precise control over how the agent loop behaves
- Your use case is simple and synchronous (one question, one answer)
FAQ
Is Claude Managed Agents available now?
Yes, in public beta on the Claude Platform since April 8, 2026. Any developer with access to the Anthropic API can use it right away. Multi-agent coordination (an agent that creates other agents) is in research preview, available on request.
How is it different from Claude Code?
Claude Code is a local, interactive development tool that runs on your machine. Managed Agents is cloud infrastructure for deploying agents in production. The two complement each other: you can use Claude Code to build and test, then deploy through Managed Agents. Running start onboarding for managed agents in Claude API from Claude Code connects the two. If you're still deciding on your dev environment, our Claude Code vs Cursor comparison will help you choose, and our roundup of AI tools for startups in 2026 shows where Managed Agents fits in the full stack.
How much does a typical session cost?
A 10-minute agent run costs about $0.013 in session-hours ($0.08/hour × 1/6). Tokens are added on top depending on the workload. For a one-hour document-processing task, expect $0.08 in session-hours plus tokens. In the vast majority of cases, tokens account for far more of the total cost than session-hours.
Can you use your own tools with Managed Agents?
Yes. The API supports custom tools and MCP (Model Context Protocol) integration to connect external systems. The meeting-prep agent mentioned above reads calendar and contact data through custom tools, and pulls from third-party systems (CRM, note-taking tools) via MCP.
