Build real AI agents without writing code

The problem

If you cannot write code, “build an AI agent” sounds like it belongs to engineers. Every tutorial opens with a Python file and an API loop, so you close the tab and assume real agents are out of reach. You settle for typing single questions into a chat and doing the rest by hand.

Can a non-coder build real AI agents?

Yes. The thing almost nobody says out loud: you do not write agent code. An agent is a plain-English job description, not a program. You write what the job is, who does it, and what a good result looks like, and the model does the rest. In Claude Code a reusable agent is literally a markdown file, and the coordination between agents is one prompt you type in normal words.

The three roles I run

One orchestrator prompt reads the goal and fans the work out to worker agents. Each worker does one slice of the job in its own context. Then a separate reviewer agent checks the result against the goal before anything is called done. The guide you are reading was produced this way: a worker agent drafts from my notes, a second agent fact-checks it, and I only read the results. I never wrote code to make that happen.

For the reusable version where each role is saved as a file, see running a team of agents from one prompt. To see why workers also keep your main chat sharp, see handing subagents the heavy work.

The orchestrator prompt to copy

You need no files to start. The trick is to route by cost: cheap models do the mechanical work like scanning and first drafts, your strongest model plans and does the final synthesis. Paste this into Claude Code and describe your real job in the first line:

<Describe the job in one or two sentences.>

Run this as a small team, not one chat. Use as many worker
subagents as the job needs, and pick the model per role so tokens
are spent well: your strongest model plans and reviews, cheaper
models do the parallel grunt work. Plan first. Fan the work out to
the workers. Then have a separate reviewer agent check the result
against the goal and list what is missing. Loop until it passes.
Report back only the plan, the result, and the review verdict.

Every instruction there is plain English. “Use worker subagents”, “a separate reviewer agent”, “cheaper models do the grunt work” are the whole design, and you never leave prose to express any of it.

Building an agent is describing a job well, not writing code.

Do this now

1. Pick a real job you would hand a capable assistant: research a topic, draft something from your notes, or check work you already have.
2. Paste the orchestrator prompt above into Claude Code and describe that job in the first line.
3. Read only the review verdict. If the reviewer waved through something weak, sharpen what “a good result looks like” and run it again.