Lessons and outlineLesson 8 of 11

AI Concepts in PicturesLesson 8

17 min read

Agents and Computer Use

Workflow or agent, the agent loop, and AI that clicks and types.

In plain English

An agent chooses its own steps, while a workflow follows steps you set. Use a workflow when the path is known. Computer use lets an agent click and type like a person. Set limits and keep a person in charge of final actions.

Agent
An AI that works toward a goal in steps of its own choosing.
Computer use
An agent that looks at a screen, clicks and types.
Sandbox
A safe space with no access to your private files.
Picture it
An agent that uses a computer
1
Look at the screen
A screenshot
2
Decide the next click
Or typing
3
Act
Click, type, scroll
4
Check what changed
Then repeat

The word agent is used for many things. In this course, an agent is an AI that works toward a goal in steps. It decides what to do next, uses tools, checks the result and repeats. This lesson separates agents from simpler designs and shows the newest kind, agents that use a computer like a person.

Workflow or agent?

There are two designs. In a workflow, you fix the steps in code, and the AI does small jobs inside them. In an agent, the AI chooses the steps itself.

A good rule: use a workflow when you can. Use an agent only when the path really cannot be known ahead of time.

The agent loop

Try it yourself
Run an agent, one step at a time

Watch the loop, the tools and the approval.

Step through one run of a small scheduling agent. Watch the loop, the tools, the step limit and the approval.

  1. GoalBook a 30 minute meeting with Asha next week.
Step 1 of at most 10 allowed

Computer use: AI that clicks

Some agents can use a computer the way you do. They look at screenshots. They move a pointer, click and type. This lets them use software that has no tool for them. They can fill forms and move between sites.

  • Strength: it works with almost any program.
  • Weakness: it is slower than a direct tool.
  • Weakness: small mistakes pile up over many steps.
  • Risk: a web page can show text meant to trick the agent.
Picture it
Workflow and agent
Workflow
  • You decide the steps
  • Predictable and cheap
  • Good when the path is known
Agent
  • The AI decides the steps
  • Flexible but less predictable
  • Good when the path is unknown

Browser agents

A browser agent does the same inside a web browser. It can research, compare prices and fill a booking form. Treat it like a new assistant on a trial week. Watch it, set limits and review what it did.

Safe habits

  • Set a step limit and a spend limit.
  • Run it in a sandbox, a safe space with no access to your private files.
  • Ask a person before anything final, such as paying or sending.
  • Keep a record of every step, so you can see what it did.
  • Start with low-risk jobs.

An example of choosing

A team wants to answer customer emails. The steps are known: read, find the order, draft a reply, send for approval. That is a workflow. A different team wants to research a market and write a briefing. The path is unknown, and the agent must pick sources as it goes. That fits an agent, with a step limit and a person reading the result.

What to watch

  • How many steps the agent takes on a typical job.
  • How often it needs a person to step in.
  • What each run costs.
  • Where it gets stuck, so you can fix the weak point.

These four numbers tell you whether the agent is helping or only looking busy.

→Count the steps. If a job needs thirty steps, it will fail more often than one with five. Break it into smaller jobs.
⚠Never log an agent into accounts with money or sensitive data until you trust it and have limits in place.
Key takeaways
  • ✓An agent chooses its own steps. A workflow follows steps you set.
  • ✓Use a workflow when the path is known, and an agent when it is not.
  • ✓Computer use lets an agent click and type like a person. It is flexible and slow.
  • ✓Set limits, use a sandbox, keep records and ask a person before final actions.
Hands-on exercise

Workflow or agent?

  1. Write three tasks, such as sorting emails, researching a topic and booking a trip.
  2. For each, say if a workflow or an agent fits better, and why.
  3. For the agent, write a step limit and a spend limit.
  4. Mark one action that needs a person to approve.
Check yourself
Q1When should you pick a workflow over an agent?show answer

When you can write down the steps ahead of time. It is more predictable and cheaper.

Q2What is computer use?show answer

An agent that looks at the screen, clicks and types like a person.

Q3Why use a sandbox?show answer

It keeps the agent away from private files and limits the damage if it is tricked.

0 / 11 finished0%