Claude Agent SDK, the agent harness without writing the loop

You get the loop rather than a diagram of one. Tools for reading and writing files, running commands and searching the web ship with it. The context management that keeps a long-running agent inside a window is already handled, and sessions resume and fork instead of starting again. The parts that normally consume a quarter arrive as configuration: hooks run your code at lifecycle points, permissions decide which tools run without asking, subagents push a focused subtask into its own context, and MCP connects the client's systems without a bespoke adapter each time. Skills, commands and memory load from a project directory the same way they do in Claude Code, so a prompt library is a folder rather than a deployment. For a forward deployed engineer, permissions plus hooks is the part that gets you through the security conversation, because you can state in code exactly what the agent is allowed to touch, and then show the file.

It is a harness for one vendor, and the documentation does not pretend otherwise. Python and TypeScript only; any other language means running the CLI as a subprocess and parsing JSON, which works and is visibly a downgrade. The deeper cost is coupling. Hooks, permissions, subagent configuration and skills are all shaped around one provider, so a client who later mandates a second one is not swapping a client library, they are rewriting the harness you sold them. Anthropic's terms also constrain how third-party products authenticate and how they may use the branding, which is a contract question you will be asked in a procurement meeting and cannot resolve with a code change. And because the agent gets a filesystem and a shell by default, the blast radius inside a client environment is exactly as small as the permission configuration you actually wrote, not the one you meant to.

Spend an hour on it if you are building a coding or operations agent on Claude and the alternative is a month of tool loops, context compaction and approval flows. Skip it if the client has a multi-provider mandate you cannot move, or if your agent is three tool calls inside a request handler, where a direct API call is far less machinery to explain to the person who inherits it.

LangGraph if provider independence and durable state matter more than a ready-made harness.