How to work with AI agents without losing control

Working well with AI agents comes down to one skill: delegating clearly and keeping a checkpoint on the outcomes that matter. Agents are fast and tireless but literal — they do what the task says, not what you meant. The teams and individuals who get real leverage treat an agent like a capable new hire: give it well-scoped work, let it run, and review the results before they ship.

What should you delegate to an AI agent?

Start with work that's clear, bounded, and reversible: drafting, research, summarizing, triaging a backlog, first-pass code. Keep for yourself anything that needs your judgment, taste, or relationships. A good rule: if you could write a two-line brief that a competent stranger could follow, an agent can probably take a first pass.

How do you keep a human in the loop?

• Write the task as a brief, not a title — say what "done" looks like.
• Have the agent propose a plan before substantial work, and approve the approach first.
• Require a summary before anything closes; review it where the task lives.
• Cap runs with budgets so a session can't run away.

Set up guardrails once

The trick is to make the guardrails structural, not something you police by hand. In Done, an agent can't close a gated task itself — it proposes completion and waits for approval. Only a human can make a task auto-approvable, run budgets bound each session, and dependencies keep an agent from jumping ahead of blocked work. Configure it once and the system holds the line for you.

Build the habit

Working with agents is a loop: brief, delegate, review, refine. The first few times you'll over-specify and over-check. Soon you'll know which work to hand off on sight — and your list starts moving even when you're not the one moving it.