You will be able to define an AI agent and decide where a human should approve its steps.
An AI agent is a language model that keeps working towards a goal on its own, using tools that change things outside the chat. Because an agent acts, its errors become actions, and an error made early gets built into every later step. Two habits keep this in check: give the agent as little access as the job allows, and make it wait for your approval before anything risky.
Lesson 7.3 showed a single tool call: the model requests a tool once, gets the result and carries on with its answer. An agent repeats that request again and again. You give it a goal. It plans a step, calls a tool, reads the result, decides the next step and calls another tool. The loop continues until the agent decides the goal is met, or until it gets stuck.
Each step still runs on next-text prediction from Lesson 3.2, where the model predicts the most likely next text. Inside an agent, the predicted text is a request for the next tool.
An agent needs two things: the loop, and tools that reach beyond the conversation. Those tools might be a browser that can click and type, or access to your files, email or calendar. With them, an agent can browse websites, fill in forms, edit spreadsheets, move files and send messages. They are useful because they change things outside the chat, and that same fact is what makes them risky.
When a chat assistant makes a mistake, you get wrong text. You can read it, spot the problem and ignore it. When an agent makes a mistake, it has usually acted before you look: a booking is made, a form is submitted, an email goes to a client, a file is overwritten. Fixing a bad summary takes about a minute (a general estimate). A payment to the wrong place, a deleted folder or a message in the wrong group chat can cost much more, and some of these cannot be reversed.
Kumar (an invented example) is the office manager at a small accounting firm and hands the team's year-end lunch to an AI agent. The task is to find a restaurant near Raffles Place with room for twelve on the third of December, book it, and send everyone a calendar invite.
The agent works through these steps in order:
search for restaurants near Raffles Place open three results and read their pages check which ones take group bookings open the booking form of the best one and fill it in read the confirmation open the calendar and create an event add twelve people from the address book send
Near the start, the agent sees the date as 03/12 and reads it month first, day second, the American way. Twenty minutes later it tells Kumar the whole job is done. The booking, the invites and the confirmation email were all carried out correctly, and all of them carry the wrong date. On the day, twelve people arrive at a restaurant that has them booked for the twelfth of March.
Most agent failures follow this shape: every action worked as intended, but one wrong detail ended up in all of them. Each step of a multi-step task relies on the steps before it. Later steps inherit an early mistake, and none of them is built to go back and check it.
This is the same effect you saw within a single answer in Lesson 3.2, where whatever follows a token has to fit that token. An agent does this with actions instead of words. In a ten-step task with a misreading at step two, steps three to ten all inherit it. A longer chain gives more chances for a slip and lets each slip travel further.
The agent's final report often looks like a clean success, because from its point of view each step did what the previous step set up. Nobody, including the agent, notices the early error until a person compares the result with what they actually wanted.
Agents that read web pages or emails face an extra risk. Text can be written to trick them, such as hidden instructions on a page telling the agent to do something you never asked for. The course "AI risks: privacy, bias, deepfakes and scams" covers this in more depth.
Agents still have a place. On tedious multi-step work, such as gathering prices from several suppliers' websites into one table, they save real time. The aim is to set the job up so mistakes are caught before they become actions you cannot undo.
The first habit is least access: give the agent only the access the task calls for. An agent that only needs to read your calendar to find a free slot should not be able to send email. An agent collecting supplier prices has no need for your company card.
The second habit is approval before risky steps: the agent waits for your go-ahead before spending money, sending a message to anyone, or deleting or overwriting anything. Many agent tools have a setting for this, and some ask for approval by default. Place approval points where a mistake would stop being text and become an action. For Kumar, one approval step before the booking, showing the restaurant, the date and the number of people, would have caught the error in seconds.
Think about a task in your own job that has several steps, and picture where in that chain you would want the same pause.
Pick one multi-step task from your job. Mark which steps you would let an agent do alone, which need your approval and which you would never hand over.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).