Claude Managed Agents Just Hit Public Beta. Here Is What It Does, How To Set It Up, And Who It Impacts.

Anthropic just moved Claude Managed Agents into public beta. This is the version everyone gets, not the early-access program. If you have a Claude API key, you can build an agent today and Anthropic runs it for you. No infrastructure, no orchestration loop, no sandbox to babysit.

This post is the practical breakdown. What it actually does, how to set it up step by step, who it helps, and who has to scramble. We are going to skip the marketing language and give you what you need to make a call this week.

What Claude Managed Agents actually is

Think of it as Anthropic running the entire “AI agent” stack for you in their cloud, while your code just describes what the agent should do.

You define one thing: an agent. That includes the model, the system prompt, the tools it can use, and the guardrails. Anthropic handles everything else. The container the agent runs in. The session that tracks what it has done so far. The retries when something fails. The auth tokens for every tool it calls. The audit trail of every decision it made.

Before this, building a production agent meant hiring engineers to wire up a half dozen separate systems. Now you describe the agent and hit start.

What is new today (April 8, 2026)

• Public beta is live for everyone. Enabled by default on all API accounts. The beta header is managed-agents-2026-04-01.
• New Claude CLI for building and shipping agents from your terminal, alongside the Claude Console and Claude Code.
• Define agents in plain English or YAML. Non-engineers can ship an agent without writing a single loop.
• Pricing is published. Standard token rates plus $0.08 per session hour of active runtime, billed in milliseconds. Idle time is free.
• Web search add-on: $10 per 1,000 searches if your agent needs to read the live internet.
• Rate limits: 60 session creates per minute and 600 session reads per minute, per organization. Plenty for most teams.
• Three research previews gated behind a waitlist: Outcomes (self-evaluating loops), Multi-agent coordination (lead agent spawns sub-agents), and Memory (cross-session state).

The four core ideas you need to know

Anthropic split the runtime into four pieces. Once you understand how they fit together, the whole platform clicks.

1. Agent. The blueprint. Model, prompt, tools, MCP servers, skills. You create it once and reuse it across sessions.
2. Environment. The container template the agent runs inside. Has Python, Node, Go, your custom files, and your network rules baked in.
3. Session. One running instance of an agent inside an environment. Sessions can live for hours. They survive disconnects. Their full history is stored on Anthropic’s servers.
4. Events. Everything that happens in a session streams back to your app over a live connection. Every model turn, every tool call, every result. You can inspect them all.

How to set it up in 10 steps

This is the actual onboarding flow. You can have a working agent in under an hour.

1. Get your Claude API key. Go to claude.com/settings/keys and create one. Free to start.
2. Install the new Claude CLI. Anthropic shipped it today. The CLI is the fastest path to a working agent if you are technical. The Claude Console is the fastest path if you are not.
3. Pick your first job. Start small. A document summarizer. A code review agent. An invoice processor. Something with a clear input and a clear output. Do not try to build the everything agent on day one.
4. Define the agent. Give it a name, pick the model, write the system prompt, list the tools it can use. You can do this in plain English, in YAML, or in code. Start in the Console if you have never done this before.
5. Pick an environment. Anthropic provides standard templates with Python, Node, Go, common command line tools, and web access. If you need a custom one, you can extend a template.
6. Wire in your tools. The built in tools cover bash, file editing, web search, web fetch, and any MCP server you connect. Auth tokens go in Anthropic’s vault and are never exposed to the running container.
7. Start a test session. Send the agent its first task. The Console shows you everything it does in real time. You can also stream events directly into your own app.
8. Watch what it does. Read the trace. See which tools it called, what it tried, where it got stuck. The audit trail is the most underrated feature here. You learn fast.
9. Tighten the guardrails. Restrict which tools the agent can use. Set spending limits on the session. Define which files it can touch. Give it the smallest set of permissions that lets it do the job.
10. Ship it to one team. Do not roll it out company wide on day one. Pick one team that owns the workflow. Let them use it for a week. Listen to what breaks. Then roll wider.

The numbers that matter

• $0.08 per session hour is the headline price. A session that runs 30 minutes a day for a month costs you about $1.20.
• 10 times faster time to production according to Anthropic and launch partners.
• Up to 10 points higher task success rate on structured work compared to a hand built loop.
• 60 percent faster first response time and 90 percent faster worst case latency than the previous Anthropic harness.
• Rakuten stood up specialist agents for product, sales, marketing, finance, and HR in one week per agent.
• Sentry shipped its automatic code fix agent in weeks instead of months.
• Six named launch partners so far: Notion, Rakuten, Asana, Sentry, Atlassian, and Vibecode.

Who this impacts (and how much)

This is a runtime layer, which means it changes the calculus for almost everyone in the AI build stack. Here is the breakdown.

1. Enterprise IT and platform teams. (Big winners.) Sandboxing, identity, observability, audit, and compliance are now Anthropic’s problem, not yours. The internal infrastructure project that was going to take six months and three engineers can probably be skipped. The proof: Rakuten put an agent in every department in five weeks total.
2. Startups and small teams. (Even bigger winners.) You can ship a real production agent without an infra hire. At eight cents an hour, the runtime cost is rounding error compared to a single engineering salary. Speed of iteration goes up by 10x.
3. No code builders and operators. (New entrants.) The natural language and YAML definition path means you do not need to write a single agent loop. This is the first time Anthropic has shipped a surface that a non engineer can actually use. Expect a wave of operator built agents this quarter.
4. Existing Claude API customers. (Optional upgrade, no migration.) The Messages API still exists for fine grained control. Managed Agents is the new default for anything async or long running. You are not forced to switch, but most new agent work should start here.
5. AWS Bedrock Agents. (Forced to react.) Anthropic just launched a competing managed agent runtime that runs Claude better than AWS does. AWS will need to either match the developer experience or get squeezed.
6. OpenAI Assistants and Responses API. (Direct competitor pressure.) Anthropic’s pricing, partner list, and architecture are now public. OpenAI will respond. Watch for an Assistants v2 within weeks.
7. LangChain and LangGraph. (Commoditized.) The agent loop layer is the thing they specialize in. Anthropic just gave it away for free as a managed service. Their value moves up the stack toward observability and tooling.
8. CrewAI and other multi-agent frameworks. (Squeezed.) Multi agent coordination is in research preview today and will land in beta within weeks. Their differentiation just got a deadline.
9. Sandbox as a service providers (E2B, Modal, Daytona). (Threatened on Claude workloads.) If your Claude agent already runs in Anthropic’s sandbox for free, you are not paying a third party for one. They keep their value for non Claude workloads.
10. Anyone with a half built agent in production. Audit your stack this week. If you built sandboxing, sessions, retries, and auth from scratch, calculate what your monthly engineering cost is for that infrastructure. Compare to eight cents an hour. The math is going to embarrass you.

What is still in research preview

Three features are not in the public beta yet. They are in a separate gated preview. If you want them, you fill out the form at claude.com/form/claude-managed-agents.

• Outcomes. You define what success looks like. The agent self evaluates after every attempt and iterates until it hits the bar. This is where the 10 point task success improvement comes from.
• Multi-agent coordination. One lead agent breaks a job into smaller pieces and farms them out to other agents in parallel. You stop writing the orchestration code yourself.
• Memory. The agent remembers things across sessions. Today you have to pass relevant history in every time. With Memory, the agent keeps a persistent store you can read and write.

What to watch out for

This is not all good news. A few things to think about before you bet your roadmap on it.

• Vendor lock in is real. Once your agent definitions, sessions, and event history live on Anthropic’s servers, switching to another model provider is not a one line change. Plan for it.
• It is still beta. Anthropic has been very stable but this is technically not generally available yet. Do not put it on your most critical workflow on day one.
• The pricing is consumption based. An agent that runs 24 hours a day in active state costs you about $58 a month per agent at the headline rate, plus tokens, plus web search if you use it. Multiply by your fleet size before you commit.
• Auth token vault is great but not magic. If your tools need access to sensitive systems, you still need a security review. Anthropic protects the tokens but the actions the agent takes are still your responsibility.

The bottom line

Anthropic just turned the agent runtime into a commodity. What used to take a four person engineering team six months now takes one person an afternoon. The cost is eight cents an hour. The launch partners are real production deployments at companies you have heard of.

If you have been waiting for the right moment to ship your first real agent, this was it. Pick one workflow this week, define the agent in the Console tomorrow, and have a working demo by Friday. The infrastructure problem that has been blocking everyone is gone.

The new question is not “can we ship an agent?” The new question is “which workflow do we ship first?” That is a much better problem to have.