Claude Agent SDK vs Managed Agents: When to Migrate

The Claude Agent SDK and Claude Managed Agents both give Claude an agent loop with tools. The SDK runs that loop inside your own process. Managed Agents runs it for you on Anthropic’s infrastructure. That difference decides who owns state, sandboxing, recovery, and scaling.

For a primer on the hosted product, read what is Claude Managed Agents.

Scope of the Claude Agent SDK

What the SDK includes

The Claude Agent SDK (@anthropic-ai/claude-agent-sdk for TypeScript, claude-agent-sdk for Python) packages the harness behind Claude Code as a library. You get:

The agent loop: context building, model calls, tool dispatch, compaction.
Built-in tools for files, shell, and search, plus custom tools and MCP servers.
Skills, subagents, hooks, and permission modes to shape behavior.

What you do not get is anything outside your process. The harness and its tools run wherever you run the SDK: a laptop, a container, a CI job. If you want isolation, you put the whole process inside a sandbox. OpenAI’s cookbook guide on migrating from the Claude Agent SDK describes exactly this pattern: the sandbox contains the harness.

That is ideal for a single agent embedded in an application or a developer workflow. It becomes work once you need many concurrent sessions, resumable state, or multi-tenant isolation.

Claude Agent SDK vs Claude Managed Agents

	Claude Agent SDK	Claude Managed Agents
Where the harness runs	Your process	Anthropic’s infrastructure
Session state	In memory / your persistence	Hosted, durable, resumable
Sandbox	Your process or the container you put it in	Anthropic cloud sandbox or self-hosted sandbox
Crash recovery	Yours to build	Built in
Concurrency and scaling	Yours to build	Built in
Event stream for UIs	Yours to expose	SSE event stream
Billing	Tokens + your compute	Tokens + $0.08 per running session-hour
Models	Claude	Claude
Control of the loop	Full (it’s a library)	None (hosted harness)

Both run Claude only. The SDK has no platform fee and leaves operations to you. Managed Agents runs the operations for a session-hour fee. For the cost side, see Claude Managed Agents pricing.

Where teams move from the SDK

Toward OpenAI Agents SDK (cookbook pattern)

OpenAI publishes a cookbook guide for moving from the Claude Agent SDK to the OpenAI Agents SDK. Its main architectural point is moving the harness outside the sandbox and treating the sandbox as a tool, mapping CLAUDE.md to AGENTS.md, and translating skills, guardrails, approvals, and handoffs. It is a vendor switch: you trade Claude-only for OpenAI-first, and you still operate the runtime yourself. See the OpenAI Agents SDK / API comparison for how the OpenAI pieces fit together.

Toward a managed multi-model runtime

The other path keeps the managed-runtime benefits of Claude Managed Agents but drops the single-vendor constraint. A multi-model managed runtime handles sessions, sandboxes, recovery, and credentials, and treats the model and harness as configuration.

A managed runtime you can self-host with any model

In the comparison table, both columns list Claude as the only model. The SDK leaves every operational row to you, and Managed Agents runs them on Anthropic’s cloud. A third option combines the following:

	Needed
Harness	Managed for you, but readable and replaceable
Sessions, recovery, event stream	Built in
Sandbox	Your choice of provider
Models	Any provider, chosen per agent
Deployment	Your infrastructure, or hosted

OpenMA is an open-source runtime with these properties. It keeps the Claude Managed Agents API shape (agents, environments, sessions, events) and runs agents on Claude, OpenAI, DeepSeek, OpenRouter, or a custom endpoint. You can use its built-in harness, write your own, or bring Claude Code or Codex into a managed sandbox through ACP adapters.

Where the SDK sits

In the brain-and-hands split Anthropic describes for Managed Agents (see what is Claude Managed Agents), the Claude Agent SDK is the brain side: a harness for Claude. Scheduling sessions, durable state, and sandbox management come from the runtime around it. For the category view, see AI agent platform & orchestration.

Moving from the SDK to a managed runtime

Moving off the in-process SDK to a managed runtime, Anthropic’s or an open one, goes smoother with this list:

Inventory tools. Separate built-in file/shell tools (provided by the runtime) from custom tools and MCP servers you need to register.
Move instructions. Your system prompt, CLAUDE.md-style project notes, and skills become agent configuration and mounted skills.
Define the environment. Packages, networking rules, and the sandbox your tools need.
Move secrets out of the process. Put API tokens in a vault so the runtime injects them into outbound requests and the sandbox never holds them.
Replace your loop with sessions. Create a session per task or conversation, send user.message events, and stream events back.
Map approvals. Permission prompts become tool permission policies and user.tool_confirmation events.
Handle custom tools. Client-executed tools pause the session with requires_action. Reply with user.custom_tool_result.
Decide on models. If you are moving to an open runtime, decide which agents stay on Claude and which move to other providers.
Measure cost per completed task before and after.

OpenMA-specific steps are in the OpenMA docs, and teams coming from the hosted product can follow migrating from Claude Managed Agents.

When to stay on the SDK only

Stay on the Claude Agent SDK when:

You run one agent, or a few, inside an application you already operate.
Sessions are short and losing one on a crash is acceptable.
Claude is your only model and you want full control of the loop.
You do not need a multi-tenant API or an operator console.

Move to a managed runtime when you need durable sessions, isolated sandboxes per task, many concurrent users, or credentials kept out of the agent’s reach. If you also need more than one model vendor or your own infrastructure, compare the Claude Managed Agents alternative options.