Explainer Plain English

Claude Managed Agents: Delegation, Explained

Anthropic built the control tower — sandboxing, checkpoints, a grader model that fails your agent until it passes. I read the whole spec like an SRE. Here's the paper verdict.

String musicians performing together in an orchestra — many players, one score
Photo by Samuel Sianipar via Unsplash
The receipts
  • Claude Managed Agents is Anthropic's suite of composable APIs for cloud-hosted agents — public beta on the Claude Platform since April 8, 2026.
  • Beta pricing: standard token rates plus $0.08 per session-hour of active runtime. Beta rates can change.
  • Outcomes (research preview) has a separate grader model score your agent's work against a rubric you write, and the agent iterates until it passes. Anthropic's internal testing claims up to 10 points better task success — vendor-reported.
  • General context on why guardrails matter: AgentCIBench (June 2026) found 11 of 15 computer-use agents leaked private info in over 50% of scenarios, averaging 67.9% leakage. Not a test of Managed Agents specifically.
Short answer

Claude Managed Agents is Anthropic's public-beta suite of composable APIs for running cloud-hosted AI agents on the Claude Platform, launched April 8, 2026. It bundles sandboxing, checkpointing, credential management, scoped permissions, cross-session memory, and tracing as managed infrastructure, priced at standard token rates plus $0.08 per active session-hour during beta.

What is Claude Managed Agents?

Claude Managed Agents is Anthropic’s suite of composable APIs for cloud-hosted agents, in public beta on the Claude Platform since April 8, 2026. You get sandboxing, checkpointing, credential management, scoped permissions, memory across sessions, and integrated tracing as managed infrastructure. Pricing during beta: standard token rates plus $0.08 per session-hour of active runtime.

I got my start in Ops — United GNOC in Chicago, rapid-response. So when Anthropic ships a product that is basically “we run the control tower, you fly the plane,” I read the spec the way I used to read runbooks.

Straight up: I have not run Managed Agents in production. This is the paper verdict… and the paper is interesting.

Why is Anthropic selling infrastructure instead of intelligence?

Because they think infrastructure is the actual bottleneck. At Code with Claude on May 6, 2026, Anthropic framed production agents as an infrastructure problem, not an intelligence problem — the same event where Dario Amodei said Q1 2026 annualized usage grew 80x against a planned 10x.

That framing matches everything I see building AI agents for clients. The model is rarely why the pipeline dies. It’s the boring parts — credentials, checkpoints, permissions, knowing what the agent did at 3am. Every shop rebuilds that plumbing from scratch.

Anthropic is selling the plumbing. That’s a real product.

How does the Outcomes grader loop work?

Outcomes is a research preview inside the beta: you write a rubric describing success, and a separate grader model evaluates the agent’s output against your criteria, pinpoints the gaps, and the agent iterates until it passes.

One model to build. One model to grade. Repeat until it clears the bar.

Anthropic’s only published quality number: internal testing showed up to 10 points better outcome task success than standard prompting loops on structured file generation. That’s vendor-reported, on their own eval. Respect the mechanism, hold the number loosely.

Multiagent orchestration — also a research preview — is the other half: a lead agent splits the job and delegates to specialist subagents with distinct models, prompts, and tools, working in parallel on a shared filesystem. Netflix has deployed it for its platform team, per 9to5Mac. Same shape as the subagent pattern in Claude Code and Cowork, which I run daily — but hosted, metered, and graded.

Should you trust cloud agents with real access?

Not blindly. AgentCIBench (arXiv, June 2026) tested 15 frontier and open-weight computer-use agents in synthetic app workspaces — general agents, not Managed Agents — and 11 of 15 leaked scenario-specified private info in over 50% of scenarios. Average leakage: 67.9%.

Capable but careless. That’s the whole category right now, and it’s exactly why scoped permissions and sandboxing being default infrastructure matters more than any benchmark.

Is Claude Managed Agents worth using right now?

The delegation model is the right architecture — it’s how I’d wire my own agency stack, and honestly I want to route iExcel client work through exactly this shape once the previews harden. But wanting is not shipping, and I don’t fake receipts.

So here’s the honest read: the beta infrastructure is buyable today. The research previews — Outcomes, orchestration — are the exciting part and the unproven part. Price the $0.08/session-hour meter into any agent stack you’re planning for 2026, and delegate work you can grade before work you must trust.

The control tower is built. Now we see who can actually land planes.

#TheAIMogul

Bottom lineManaged Agents is Anthropic selling the part of agents nobody wants to build — the control tower, not the pilot. The infrastructure thesis is correct; the research-preview features aren't load-bearing yet, so delegate work you can grade, not work you must trust.

Frequently asked

What is Claude Managed Agents?
It's Anthropic's suite of composable APIs for cloud-hosted agents, in public beta on the Claude Platform since April 8, 2026. Instead of you building sandboxing, checkpointing, credential management, scoped permissions, memory, and tracing yourself, Anthropic runs that infrastructure and you compose agents on top of it through the API.
How much does Claude Managed Agents cost?
During the public beta you pay standard Claude Platform token rates plus $0.08 per session-hour of active runtime. That runtime meter is the new line item — a long-running agent accrues session-hours on top of its tokens. It's beta pricing, so expect the rate to change before general availability.
How do Outcomes rubrics work in Claude Managed Agents?
Outcomes is a research preview where you write a rubric describing what success looks like. A separate grader model evaluates the agent's output against those criteria, pinpoints what needs to change, and the agent iterates until it passes. Anthropic's internal testing reported up to 10 points better outcome task success versus standard prompting loops on structured file generation — a vendor-reported number.
What is multiagent orchestration in Managed Agents?
Also a research preview: a lead agent breaks a job into pieces and delegates each piece to specialist subagents — each with its own model, prompt, and tools — running in parallel with shared filesystem access. Anthropic announced it at Code with Claude on May 6, 2026, and 9to5Mac reports Netflix has deployed it for its platform team.
Are cloud agents safe to trust with sensitive data?
Be careful with the category. AgentCIBench (arXiv, June 2026) tested 15 frontier and open-weight computer-use agents in synthetic workspaces — not Managed Agents specifically — and 11 of 15 leaked scenario-specified private info in over 50% of scenarios, averaging 67.9% leakage. Scoped permissions and sandboxing exist precisely because agents are capable but careless.