Describe the agent you want, what it should do, and which tools or APIs it needs, and get real, runnable starter code, not pseudocode. Pick a framework for a working agent loop with tool stubs, or pick IDE skill prompt to get a base prompt you paste into Claude Code, Cursor, or Windsurf that builds the agent inside your own repository.
Each of these describes an agent at the level of detail the tool works best with: the job, the tools, and the one rule that matters.
Instead of code, it writes an instruction block you paste into an AI coding IDE such as Claude Code, Cursor, or Windsurf. That block acts as a reusable skill: it tells your editor to inspect your actual repository first, match your existing conventions and HTTP client, create a named set of files, and build the agent loop to a defined spec, with checks it can run to prove the result works. You get an agent that fits your codebase rather than a generic file you have to adapt.
Use the code when you want something runnable in isolation to try an idea. Use the skill prompt when the agent has to live inside an existing project, since your IDE can see your real dependencies, config, and conventions and build to them.
It generates a working starter scaffold. The agent loop, tool definitions, and structure are real and runnable, but you'll want to wire in your actual APIs, auth, and error handling before shipping.
Claude Agent SDK, OpenAI Agents SDK, LangGraph, and a plain dependency-free Python loop, plus the IDE skill prompt mode. Pick whichever matches your stack.
Yes, and you should. Mention the tools or APIs you want, for example 'a Slack tool and a calendar tool', and they'll be scaffolded as stub functions with correctly-shaped return values so the loop runs end to end straight away.
So the loop runs on first execution. A stub that returns a correctly-shaped fake response lets you watch the whole agent work before you have credentials for anything, then you swap one body at a time for the real call.