Last updated: 2026-10-09
Agents in ChatGPT, Claude, Copilot and Gemini can now act across your browser, files and email, sometimes on a schedule. Eight lessons on picking the right jobs, briefing them well, and keeping the limits, approvals and off switch in your hands.
Self-paced. Honour-based. Not accredited.
Agents are fast, literal and tireless, and their mistakes happen in your real accounts. This course gives you a simple method: choose checkable jobs, brief them clearly, give the least access, and keep approvals and an off switch.
From what an agent is to a written plan for your first delegated job.
What agents do, which jobs suit them, and how to brief them so they do the right thing.
Least access, approval gates, scheduled tasks, budgets and the kill switch.
Prompt injection, risky plugins, and when a chat, a filter or you are the better choice.
Start with what makes an agent different from a chatbot, and what no agent can promise.
New to AI at work? AI at Work covers the everyday office tasks first. For the engineering behind agents (RAG, MCP, evals and multi-agent systems), continue to Beyond the Prompt. Running your own agent? Our Hermes Agent security guide and agent security round-up go further.
Course progress, exam results and certificate details are stored in this browser on this device. They may be removed if you clear browser data.
A chatbot answers. An agent acts: it plans steps, uses tools such as your browser, files and email, checks the result and carries on. That makes it more useful, and it means its mistakes happen in the real world.
The best first jobs for an agent are boring, repeatable, easy to check and easy to undo. The worst are anything involving money, irreversible actions or other people's private information.
An agent follows your brief literally and fills gaps with guesses. A good brief states the goal, the context, the limits, what done looks like, and when to stop and ask you.
Whatever an agent can reach, it can get wrong, and so can anyone who tricks it. Give it the least access the job needs, and make it show you before anything leaves, gets paid for or gets deleted.
Agents that run on a timetable are the most useful kind and the easiest to forget. Keep a list, check the results, and switch off anything you no longer read.
Agents use far more computing than a chat: every step reads pages and writes notes. On a subscription that means hitting limits; on pay-as-you-go it means real bills. Set a budget and know how to stop it.
Agents read web pages, emails and documents, and they can't reliably tell your instructions from instructions hidden in that content. That is prompt injection, and it is the main way agents get turned against their owners.
Sometimes the right answer is a chat, a simple automation or doing it yourself. Finish by choosing one job to delegate, writing its brief and limits, and deciding how you'll check it.