Building & appsClaude

🚦 Set up your first Claude agent (human-in-the-loop)

An agent is AI that doesn't just answer — it does: multi-step work using tools, files, and the web. The difference between useful and scary is one design pattern: human-in-the-loop. Here's your first agent, built with approval gates from day one.

Pick a first job with training wheels

Good first agent jobs are multi-step, repetitive, and reviewable before anything happens:

Research 5 suppliers and draft comparison notes
Read this folder of invoices and build a summary sheet
Draft replies to these 10 emails (drafts — not sends)

Bad first jobs: anything that sends, deletes, posts, or pays. Those come later, if ever.

The agent instruction template

You are my [job] agent. Your task: [the multi-step job, concretely]. OPERATING RULES: 1. Start by showing me your plan as numbered steps. Wait for my OK. 2. Work through the steps, showing your output as you go. 3. STOP and ask before: anything irreversible, anything leaving a draft state, anything touching data outside [scope], or anything where you're less than sure. 4. If a step fails, tell me what happened — never improvise around an error silently. 5. End with: what you did, what you couldn't do, and what needs my review.

That's the human-in-the-loop pattern: plan approval up front, gates at consequences, honest reporting at the end. It runs in Claude with tools/connectors enabled, in Claude Code for file work, or in Cowork — same template everywhere.

Widen the leash slowly

Weeks 1-2: agent proposes, you approve every step. Tedious on purpose — you're learning its failure modes.
Weeks 3-4: pre-approve the routine steps, keep gates on consequences.
After a month: the agent runs its job start to finish, and the gates remain permanently on: sending, deleting, paying, publishing.

Scope its access like a new hire's. Give it the folder it needs, not your whole drive; the account it needs, not your passwords. Trust is built per-task, and some permissions (money, mass communication) should stay human forever.

Keep an agent log. One note: what you asked, what it did, where it stumbled. Two weeks of logs tells you exactly which second agent to build — and how it fits your bigger assistant setup.

Want this done for you?

We set up these exact systems for businesses like yours. One call, honest scope, no pressure.

Start your project