Give AI a project. Stay involved.

Learn how to delegate work to an agent, set permissions, and review its results.

What an agent adds

A chat produces a response. An agent can use tools to read files, research, create documents, run code, and work through multiple steps. What it can actually do depends on the application, enabled tools, and access you grant.

Write a brief someone could work from

Describe the finished result, the audience, the allowed inputs, and an example of success. Ask for a small sample before a long run. This gives you a chance to catch a misunderstanding while it is cheap to fix.

Make permissions concrete

Use a copy or a new folder. Limit access to the files the project needs. Require approval before messages, purchases, publishing, or destructive changes. Text the agent reads from websites or documents can contain malicious instructions; treat those materials as inputs, not authority. Approval settings reduce risk but are not a guarantee.

Review at milestones

Ask for an inventory, a sample, a draft, and a final deliverable as separate stages. At each stage, inspect actual files and sources. If a check failed, the final report should say so. A confident completion message is not verification.

Try a project worth caring about

Make a neighbourhood resource guide from public sources, build a small tool for a hobby, or turn your own notes into a workshop. Keep a record of what you asked, what changed, what failed, and what you checked. Publish measured results only when you have those records.

Try it

• Build a small tool you wish existed.
• Bring order to a messy project folder.
• Take on a project you kept putting off.

Go deeper

Follow a learning path or try an interactive lab.

Related Power of AI pages

Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.