# Agents SDK (/academy/agents-sdk)



An Agents [SDK](/academy/sdk) is a prefab workshop: the benches are already bolted down. You walk in and hang the job on the wall.

You are not training a model. You are assembling a [harness](/academy/harness) — the loop, [tool calling](/academy/tool-calling), a place to remember step one when you hit step four, [guardrails](/academy/guardrails) that do not trust "please don't delete." OpenAI, Anthropic, the Vercel AI SDK: different boxes, same shop.

Building that from scratch is days of plumbing before the idea. The kit hands you:

1. **The loop.** Step, check, step again. That is what makes a [worker](/academy/worker), not a chatbot.
2. **The wiring.** Call an [API](/academy/api), run a command, send the mail.
3. **The clipboard.** Enough memory that the run has a plot.
4. **The fences.** Ask before refunds. Stay on this task.

**You may not need one**

If all you need is "scaffold this [stack](/academy/stack)," install the plugin or point an existing chat app at `npx create-hackhaton-space-stack@latest mcp`. The [agent](/academy/agent) is already housed. The tools are already a menu. Reach for an SDK when the worker *is* the product — a support bot that is your app, a voice line, a research runner you ship.

The kit is a frame, not a trustworthy employee. Tightening the rails, picking the tools, watching the first dozen runs: still yours.

**What this unlocks**

The machinery stops being the project. The mission starts. Teams that look like they "built an agent" usually opened a kit, described a job, and plugged in their windows.

The kit handles the shop. You bring the work order.