Claude Managed Agents is Anthropic’s hosted agent: Claude, a harness, and a sandbox per session, billed at $0.08 per session-hour while it works. It’s a good product with a short list of hard limits, and each limit points at a different alternative.

Claude only? Look at Amazon Bedrock AgentCore, LangSmith Deployment, or Cadenya. Need Zero Data Retention or HIPAA? AgentCore. Want OpenAI’s version of the same idea? The Agents API. Want your own API as the tools, with any model? Cadenya.

Every fact below comes from the vendors’ own docs and pricing pages, checked on September 28, 2026.

What Claude Managed Agents does well

Start with why people pick it. Anthropic describes it as a “pre-built, configurable agent harness that runs in managed infrastructure”, and it delivers on that:

• A sandbox container per session. Claude can run bash, read and write files, fetch pages, and search the web without you building any of it.
• Human-in-the-loop approvals that wait. Set a tool to always_ask and “the session waits indefinitely for a response” until you send a tool confirmation (a user.tool_confirmation event).
• You pay for work, not waiting. Runtime is $0.08 per session-hour, and time spent idle doesn’t count (the docs state no session idle timeout). Tokens bill at the usual rates (Claude Sonnet 5 is $2 and $10 per million input and output tokens).
• The rest of the kit. Up to 20 MCP servers per agent, memory stores, multi-agent sessions, SDKs in seven languages, and self-hosted sandbox environments on 11 providers (the default environment is Anthropic’s cloud).

The limits in its own docs

Five limits come straight from Anthropic’s pages:

1. Claude only. An agent’s model is “the Claude model that powers the agent”, 4.5 or later.
2. One model per session. “The agent’s model configuration, including its inference_geo pin, also can’t change mid-session.”
3. No ZDR or HIPAA. Managed Agents “is not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement (BAA) coverage”, and "us" is the only workspace geo.
4. Beta. “Claude Managed Agents is in beta”, and every request carries the managed-agents-2026-04-01 beta header. Anthropic’s pages give no general availability date.
5. The loop stays at Anthropic. Even with a self-hosted sandbox, “Tool inputs and outputs still flow to Anthropic’s control plane (where Claude runs).”

And one gap that matters if your agent should act through your product’s API: there’s no OpenAPI import. Your API comes in as MCP servers or as custom tools that your own code executes.

The alternatives

OpenAI’s Agents API

The closest in shape: a managed harness (OpenAI’s Codex harness) with sessions, compaction, and recovery. You can change the model between turns of a session, which Managed Agents can’t. But it’s also in beta, it runs OpenAI models, it “supports data residency only in the United States and does not support Zero Data Retention”, and it has no approval step of its own (OpenAI’s Agents SDK has one, in your code). Pricing is model tokens, tools, and $0.03 per 20 minutes for a 1 GB sandbox.

Amazon Bedrock AgentCore

The answer to “any model” and “compliance”. AgentCore runs your agent in its own microVM per session, with any framework and any model, and AWS lists it as HIPAA eligible, FedRAMP, SOC 2, and ISO. You bring the loop, though: AgentCore hosts the code you write. And its sessions are short by design: up to 8 hours on microVMs, 15 minutes idle by default, and the built-in approval hook waits at most 900 seconds before it denies. Runtime bills from $0.0895 per vCPU-hour, per second, with no minimum. AgentCore alternatives covers it in depth.

LangSmith Deployment

The answer if you want to write the agent and have it hosted. It runs LangGraph and, through a wrapper, other frameworks, in LangChain’s cloud in the US or the EU. LangGraph interrupts wait “indefinitely until you resume execution”, like Managed Agents’ always_ask. Plus costs $39 a seat a month plus metered runtime, and self-hosting is an Enterprise add-on. LangSmith Deployment pricing and alternatives breaks down the bill.

Cadenya

The answer if your agent acts through your own API and you want any model with nothing to run. Cadenya turns your OpenAPI document into tools, runs the loop on any provider’s model, and puts the chat in your app with a React widget. It’s free for 1,000 loops a month and $49 for 50,000 (a loop is one LLM request). What it offers that Managed Agents doesn’t is at the end.

Your own loop on a durable engine

The answer if you want every piece in your hands. Run the loop in your code with any model SDK and make it durable with Temporal, Restate, DBOS, or Inngest. Each one can hold an agent through a crash and a long approval, and each asks you to run something different: Restate vs Temporal, Inngest vs Temporal, and DBOS vs Temporal compare them.

Side by side

	Claude Managed Agents	OpenAI Agents API	Amazon Bedrock AgentCore	LangSmith Deployment	Cadenya
Models	Claude only	OpenAI	Any	Any your code calls	Any provider you connect
Model change inside a session	No	Between turns	Yes, in the harness	Your code decides	Per objective, through variations
Approval wait	Indefinite	No approval step	Up to 900 seconds	Indefinite	Until answered, within the idle timeout you set
Your OpenAPI as tools	No (MCP or custom tools)	No (MCP or functions)	Yes, through Gateway	Your code	Yes
Sandbox for generated code	Yes	Yes	Yes	Sold separately	Pairs with E2B or another sandbox
ZDR or HIPAA	Neither	No ZDR	HIPAA eligible	Not stated	Not stated
Status	Beta	Beta	GA harness	Some features in beta	Live, SDKs at 1.7.0

When to stay on Claude Managed Agents

Stay if Claude is the model you want, the agent’s work happens in files and a shell, and approvals can take days. Nothing else on this list gives you all three without building part of it yourself.

What does Cadenya offer that Claude Managed Agents doesn’t?

Managed Agents gives Claude a shell and a sandbox. Cadenya gives any model your product’s API, runs the loop, and puts the agent in your app. Here’s what that covers:

1. Any model, from any provider. Managed Agents runs Claude 4.5 and later. On Cadenya, the model is a setting on a variation: any model on OpenRouter, Claude included, or your own endpoint that speaks the OpenAI chat format, billed by your provider on your own keys.
2. Model routing between variations. Weighted variations send more traffic to whichever model or prompt users rate higher, and a variationId pins one configuration for a regression test or a model comparison.
3. Your OpenAPI document as the tools. Managed Agents has no OpenAPI import, so your API comes in as MCP servers or custom tools your own code runs. A Cadenya tool set turns every operation in an OpenAPI 3 document into a tool, calls your API over HTTP, re-checks the document each hour, and loads tools on demand. MCP servers and plain HTTP endpoints work too.
4. Built for agents inside your product. The agent sits in your app as a React chat widget, and your backend mints a session per user. That session can carry the user’s short-lived token as a secret that overrides the shared API credential, so your API sees the signed-in user on each tool call, and pinned parameters fix the IDs the model can’t change.
5. Human-in-the-loop approvals with a screen. Like always_ask, a tool marked for approval pauses the objective on toolApprovalRequested. Cadenya adds the places to answer it: the API, a webhook handler that posts to Slack, or a person in the React chat widget. A denial can carry a memo that steers the agent.
6. An idle timeout you set. Each variation sets how long an objective may sit with no activity before it closes, so abandoned conversations don’t pile up. Managed Agents documents no idle timeout for a session.
7. No beta header. Managed Agents is in beta, and every request carries the managed-agents-2026-04-01 header. Cadenya’s API needs no beta header, its TypeScript and Python SDKs reached 1.0 on August 15, 2026, and 1.7.0 shipped on September 28.
8. A price per model call, not per session-hour. Managed Agents’ pricing is session-hours, at $0.08 each while the session runs, plus Claude tokens. Cadenya bills per loop, one LLM request: free for 1,000 a month, $49 a month for 50,000, then $0.0015 each, and your provider bills the tokens.

The tool sets guide shows how your API becomes an agent’s tools, and the free plan needs no card.