On April 8, 2026, Anthropic launched Claude Managed Agents in public beta on the Claude Platform — a composable API suite for building and deploying cloud-hosted AI agents.
It handles the infrastructure work most teams spend months on: sandboxed execution, checkpointing, credential management, scoped permissions, and session tracing — all managed by Anthropic.
Internal testing shows up to 10-point improvement in task success over standard prompting loops. Early adopters including Notion, Asana, and Rakuten are already shipping agents in days.
Pricing: standard Claude Platform token rates + $0.08 per session-hour. Available now at the Claude Console.
Claude Managed Agents is Anthropic's fully managed infrastructure for building and deploying AI agents at production scale. Launched in public beta on April 8, 2026, it provides sandboxed code execution, long-running autonomous sessions, multi-agent coordination, and built-in session tracing — so teams can go from prototype to production in days rather than months. It is available on the Claude Platform with pricing at $0.08 per session-hour plus standard token rates.
What Problem Does It Actually Solve?
If you've tried shipping a production AI agent, you know the real work isn't the model. The model is the easy part. The hard part is everything around it - secure code execution that doesn't let the agent trash your servers, checkpointing so a 3-hour task doesn't vanish if a connection drops, credential management so the agent can actually touch real systems without full admin access, and tracing granular enough that you can debug what went wrong at step 47 of an 80-step run.
That infrastructure work typically takes months before a team ships anything users can touch. Most companies either build it themselves (slow), stitch together third-party tools (fragile), or quietly give up on anything longer than a simple one-shot API call.
"Until now, building agents meant spending development cycles on secure infrastructure, state management, permissioning, and reworking your agent loops for every model upgrade."
— Anthropic, official launch post
Managed Agents is Anthropic's answer: you define the agent's tasks, tools, and guardrails. They run it on their infrastructure. The orchestration harness handles when to call tools, how to manage context windows, and how to recover from errors automatically. You get the output — and full session traces showing every decision along the way.
Also Read:- How to access Claude Mythos?
What's Actually Included: Feature Breakdown
Two features are particularly worth flagging.
First, the self-evaluation loop in research preview: instead of writing a pass/fail check yourself, you describe what "good" looks like and Claude iterates toward it. That shifts a non-trivial amount of eval work off the developer.
Second, multi-agent coordination -agents that can spawn and direct other agents - is the capability that makes complex pipelines tractable. One orchestrator agent breaking down a large task and delegating subtasks in parallel is how you get hour-long autonomous workflows that don't require constant human handholding.
How to Deploy Your First Agent?
The API is straightforward. Four steps from zero to running agent:
# All Managed Agents requests require the beta header
curl -sS https://api.anthropic.com/v1/agents \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "anthropic-beta: managed-agents-2026-04-01" \
-H "content-type: application/json" \
-d '{
"name": "my-first-agent",
"model": "claude-sonnet-4-6",
"system": "You are a coding assistant..."
}'
Claude Code users can skip the manual setup entirely. Just ask: "start onboarding for managed agents in Claude API" and it walks you through configuration. The new CLI is the fastest path if you're starting cold.
Pricing: What It Actually Costs?
For most use cases, the session-hour fee is secondary to token costs — a long reasoning session will spend far more on tokens than on runtime. The $0.08/hour figure is competitive with what you'd pay to run equivalent container infrastructure yourself, and significantly cheaper when you factor in developer time not spent building it.
Who's Already Shipping With It
Anthropic shared a handful of early adopter stories that illustrate the range of what teams are building:
The Rakuten deployment is probably the most revealing one. Getting specialist agents across five different business functions shipped, tested, and deployed in a week each — that's not a marketing number. That's what "prototype to production in days" looks like in practice for an enterprise team.
Managed Agents vs. Building Your Own Agent Loop
| Consideration | DIY Agent Loop | Claude Managed Agents |
|---|---|---|
| Time to production | Weeks to months | Days |
| Sandboxed execution | Build yourself | Included |
| Session persistence | Custom state management | Automatic, survives disconnections |
| Model upgrade compatibility | Rework agent loop each time | Harness adapts automatically |
| Debugging / tracing | Custom logging infra | Built into Claude Console |
| Multi-agent pipelines | Complex custom orchestration | Native coordination (research preview) |
| Control over internals | Full access | Abstracted — less low-level control |
| Vendor dependency | None | Tied to Claude Platform |
The honest version of this comparison: if you have unusual infrastructure requirements, regulatory constraints that prevent cloud execution, or you need very fine-grained control over how the agent loop works, building your own is still the right call. If your goal is shipping something users can use as fast as possible, Managed Agents wins on almost every dimension.
What's Still in Research Preview (Not GA Yet)
Two features you'll see referenced throughout Anthropic's documentation come with an asterisk — they're in research preview and require separate access requests:
Multi-agent coordination — the ability for one agent to spin up and direct other specialized agents — is not yet generally available. You can request access via the Claude Platform form, but teams shouldn't build hard dependencies on GA timelines yet.
Outcome-based execution — where you define success criteria and Claude iterates until it meets them rather than running a fixed loop — is also in research preview. This is arguably the more interesting long-term feature, because it moves the developer's job from "write the loop" to "describe what done looks like." That's a fundamentally different relationship between the developer and the agent.
Is This the Right Time to Start Building?
There's a version of this story where you wait for the research preview features to go GA, let the beta stabilize, and start building in six months when everything is more locked down. That's a reasonable position if your use case specifically requires multi-agent coordination.
But for most teams, the public beta features are more than enough to ship useful agents right now. Long-running sessions with sandboxed execution and session tracing covers a huge portion of real-world agent use cases. And the teams that start building today will understand the platform — its limits, its sweet spots — months before the teams that wait.
The Rakuten and Asana deployments weren't experimental weekend projects. They were production workloads. That's the signal.
Frequently Asked Questions
Ready to build your first agent?
Claude Managed Agents is live in public beta. Head to the Claude Console or read the official docs to deploy your first autonomous agent today.
Open Claude Console →Sources
Anthropic Official Blog — Claude Managed Agents Launch · Claude Platform Docs — Managed Agents Overview · Claude Platform Docs — Quickstart Guide · The New Stack · Testing Catalog
Join the conversation