Back to catalog
AI at workIntermediate

AI Agents and Routine Automation

An agent does not answer — it acts. What it is made of, what to let it touch, and why the cost of a mistake goes up

AI Agents and Routine Automation

About this course

This course starts where the chat window ends. A chat returns text and you decide what to do with it. An agent gets a goal, picks its own steps, calls tools on your behalf and stops when it decides it is done. That is not a smarter model — it is the same model with hands, and a mistake now lands in your mail, your files and your accounts instead of on screen.

We take the agent apart: goal, tools, memory, loop, stop condition. We connect tools and connectors, look at what a tool-connection protocol is for, and work out which permissions to hand over — starting with read-only. We design confirmation points, bound the loop so it cannot spin, read the step log, and deal with the awkward case where the agent almost got it right.

Two lessons cover risk you cannot wave away: prompt injection through the data the agent reads, and what an agent sees and sends inside someone else's services. The course ends honestly — with the tasks where an agent is the wrong answer and a plain rule, script or schedule is cheaper, more predictable and safer.

No programming required. Nothing here depends on a model version or a particular product: products change, the principles do not.

Course curriculum

  1. From an answer to an action

    What actually changes when the model stops answering and starts doing, and which five parts every agent is built from.

    • 1.An agent is not a smarter chat12 min
    • 2.What an agent is made of14 min

  2. Tools, connectors and permissions

    How a model gets hands, what a tool-connection protocol is for, which permissions to grant, and which tasks to keep away from an agent.

    • 1.Where the hands come from: tools and connectors13 min
    • 2.Permissions: read-only is a good start13 min
    • 3.What to hand over, and what to keep12 min

  3. The loop under control

    Confirmation points, limits that stop a spinning agent, and how to check a run by its log instead of its summary.

    • 1.Confirmation and stop points12 min
    • 2.Why an agent goes round in circles14 min
    • 3.Checking the work: the log, reproducibility and 'almost'13 min

  4. The trust boundary

    The risk that belongs to agents alone: instructions hidden in the data they read, and what an agent sees, sends and signs inside someone else's services.

    • 1.Prompt injection: instructions hidden in the data16 min
    • 2.The agent inside someone else's services13 min

  5. Routine automation — and when not to automate with an agent

    Scenarios that pay off without any programming, how to roll one out, and the tasks where a rule, a formula or a schedule beats an agent.

    • 1.Simple automations that actually pay off15 min
    • 2.When a plain script or a schedule is cheaper14 min