What is an AI agent
What it is
An AI model (also called LLM, “language model”) is the program that reads text and writes text, like what’s behind ChatGPT or Claude. A tool is something AI can use to act in the real world: read a spreadsheet, send a message, check a calendar. An AI agent is a model that takes an objective and uses tools to execute steps until it reaches a result. The key difference: the chat answers, the agent executes.
Why learn
Confusing answering with executing underestimates the risk. A wrong response in chat is something you read and discard. A wrong action by an agent already happened: the message was sent, the request was changed. Knowing what an agent is is the first step to decide what it can and cannot do on its own.
Key concepts
Objective: understand the difference between a request to respond and a request to execute.
Paste into ChatGPT or Claude:
“Here are two requests. Tell me which one is RESPONDING and which one is EXECUTING, and what tool the AI would need for the second:
1) Explain how to confirm consultations via WhatsApp.
2) Confirm tomorrow’s consultations for the <your clinic / your business> via WhatsApp.”
How to verify: the AI should classify #1 as a response and #2 as execution, citing access to the schedule and WhatsApp as tools.
✓ Do this
✓ Always ask: does it only answer, or does it also act somewhere?
✗ Avoid this mistake
✗ Calling any chat conversation with a chatbot that only returns text an “agent.”
Practice before you reveal
List three tasks from your work and mark for each one whether AI would only answer or would need to execute something.
View commented answer
Example: “summarize a contract” is answering; “send a payment reminder” is executing, because it uses a tool and talks to people outside. The execution tasks are the ones that require more care with limits.
The turning point: before you executed; now you manage
What it is
Before, the path was: person → executes the process → result. With agents, it becomes: person → defines the intention → organizes the context → delegates → supervises → evaluates. Context is the set of information you give the agent so it can decide well: house rules, examples, history, and the current situation. You don’t leave the process; you just change where you stand in it.
Why learn
Someone who keeps thinking like an executor tends to delegate poorly: they throw the task over the fence and expect a miracle, or redo everything out of distrust. Thinking like a manager means aligning on what good results look like first, and checking afterwards. It’s what you already do with a new employee, now written down.
Key concepts
Objective: find out where your time will go after delegating a task.
Paste into ChatGPT or Claude:
“Today I do this task with my hands: <describe the task, e.g., answering clients about delivery deadlines>.
If an AI agent starts executing it, list what continues to be my work in four blocks: defining the intent, organizing the context, supervising, and evaluating. For each block, tell me how much time per week it should take me.”
How to verify: the answer must include all four blocks and make it clear that evaluating and correcting remain with you.
From concept to action
- Define the intention: identify the initial condition.
- Organize the context: apply the described decision.
- Delegate: verify the effect in the example.
- Supervise and assess: record the output evidence.
✓ Do this
✓ Set aside time on your calendar to review what the agent did.
✗ Avoid this mistake
✗ Delegate and disappear, assuming the agent will guess what you forgot to say.
Practice before you reveal
Think of an employee you trained. Write down what you explained on day one.
View commented answer
This is the context an agent also needs. The difference is that the employee infers what was left implicit; the agent only knows what’s written.
Intention-driven automation
What it is
Automation is making a task happen without a person performing every step. In classic automation, you design every step: if email X arrives, do Y. In intention-driven automation, you say what: the goal, what a good result looks like, the limits, and when the task is done. The AI chooses part of the route.
Why learn
Describing the intention is faster than drawing each step, but it transfers a decision to the machine. If the intention is vague, the agent carries out the wrong task efficiently. Learning how to write a good intention is the skill that pays off most in this turning point.
Key concepts
Objective: turn a vague request into a complete intent.
Paste into ChatGPT or Claude:
“My request is: <e.g., take care of the store messages>.
Rewrite it as a complete intent, in four lines:
WHAT: …
WHAT IS GOOD: …
LIMITS (what you never do): …
WHEN IT ENDS: …
If any information is missing, ask me instead of making it up.”
How to verify: the four lines appear, and the AI asked at least one question about something you didn’t say.
✓ Do this
✓ Always write all four parts, even for small tasks.
✗ Avoid this mistake
✗ Asking “take care of this for me” without saying what is good or what is prohibited.
Practice before you reveal
Take a request you would make to AI today and write the four lines: what, what is good, limits, and when it ends.
View commented answer
Example: what, confirm tomorrow’s consults; good, every patient received a message; limit, never reschedule; ends, at 5pm with the list of who didn’t respond. If you got stuck on any line, that’s where the agent would also get stuck.
The new workflow
What it is
Working with agents follows a flow: Intention → Context → Data → Execution → Evaluation → Supervision → Evolution. The source of truth is the official place where the right information comes from, like the system’s schedule—not the old spreadsheet. Autonomy is how much freedom the agent has to act without asking permission. A guardrail is a constraint that prevents prohibited actions, and a log is the record of what the agent did, so you can check afterwards.
Why learn
When something goes wrong, the flow shows where to look: vague intention, missing context, wrong data, or missing evaluation. Without this map, the common reaction is to switch tools, when the problem was in an earlier step. And the flow doesn’t end at execution: what you learn feeds back to improve the intention.
Key concepts
Objective: diagnose at which step in the flow the problem originated.
Paste into ChatGPT or Claude:
“Use this flow: 1 Intent, 2 Context, 3 Data, 4 Execution, 5 Evaluation, 6 Supervision, 7 Evolution.
This happened with the AI in my work: <e.g., it sent an old price to a client>.
Tell me which step the problem was born in, why, and the smallest possible correction in that step.”
How to verify: the AI points to a specific step, like Data, and suggests a small correction, like defining the official price table.
✓ Do this
✓ When there’s an error, first ask which step in the flow it was born from.
✗ Avoid this mistake
✗ Switching AI tools after every error without looking at intention, context, and data.
Practice before you reveal
Recall an error the AI made with you and classify which of the seven steps it was born from.
View commented answer
In most cases, the error is born in Context or Data: a rule was missing or the information was outdated. The fix is usually one more line in the instructions, not a new tool.
People + Processes + Agents
What it is
Adopting agents in a company is not just a technology issue. Without coordination, each person creates their own agent, their own spreadsheet, and their own rules—so duplicate or conflicting agents appear: two agents responding to the same customer with different information. Management then has to take care of three things at the same time: people, processes, and agents.
Why learn
A well-made agent inside a messy process makes the mess worse—only faster. In a small clinic, if the receptionist and the owner each connect their own WhatsApp robot, the patient receives two different messages. Coordination means knowing who’s responsible for each agent and which process it serves.
Key concepts
Objective: build the inventory of agents and automations in your business.
Copy this table and fill it in, one row per agent or automation that already exists:
Agent or automation ; What it does ; Process it covers ; Owner (person) ; Information source
<ex.: WhatsApp robot> ; <answers opening hours> ; <customer service> ; <Ana> ; <store website>
Then paste into ChatGPT or Claude: “Point out duplicated, conflicting, or ownerless agents in this table.”
How to verify: every row has an owner with a name, and no process appears with two agents doing the same thing.
From concept to action
- People: identify the initial condition.
- Processes: apply the described decision.
- Agents: check the effect in the example.
- Single list: record the output evidence.
✓ Do this
✓ Keep a single list of agents, each with an owner named.
✗ Avoid this mistake
✗ Letting each person set up their own automation without telling anyone.
Practice before you reveal
Do the inventory above with at least three items, including simple automations such as an automatic email response.
View commented answer
It’s common to discover forgotten automations or ones without an owner. An agent without an owner isn’t reviewed, and what isn’t reviewed ends up repeating errors without anyone noticing.
Your first delegation today
What it is
The best way to understand the turning point is to delegate a small, real task—going through the seven questions: intention, context, data, success criteria, autonomy, observation, and supervision. The Agent Sheet (Ficha do Agente) makes you answer these questions and calculates the autonomy level, from N0 (only consults) to N4 (it runs an entire process). You walk away with a ready instruction, three tests, and a checklist.
Why learn
Reading about agent management changes nothing; filling out a sheet changes things. The calculated level shows the central rule in practice: if the task involves money, irreversible action, or people from outside, the agent stays at N2—and nothing goes out without your OK. Starting small and low is the safe path to scaling up later with evidence.
Key concepts
Objective: create your first Agent Sheet in 10 minutes.
Path A (form): open https://inematds.github.io/7pa/en/?preset=clinica
Read the clinic’s ready sheet, then swap the fields for your task: <your task>.
Other ready sheets: https://inematds.github.io/7pa/galeria/en/
Path B (conversation), paste into ChatGPT or Claude:
“Be my agent management interviewer. Ask ONE question at a time, in simple language, in this order: intent, context, data, success criteria, autonomy, observation, supervision. At the end calculate the level: consult N0; draft N1; executes N3; whole process N4; if it involves money, irreversible action, or people outside, at most N2; if there’s no official source, at most N1. Deliver the sheet, the ready instruction, 3 tests, and the checklist. My task: <your task>.”
How to verify: the clinic’s sheet comes out as N2, because the message goes to people outside and can’t be undone; your sheet should bring the level with the reason in one sentence.
✓ Do this
✓ Start with a small task and accept the level the rule calculates.
✗ Avoid this mistake
✗ Raise the level by hand to “save time” before you have evidence.
Practice before you reveal
Open the gallery, choose the Agent Sheet that’s closest to your area, and note the level and the reason.
View commented answer
Examples: legal practice results in N0, because it only researches; a salon results in N1, because it only prepares drafts; a restaurant reaches N3, because it executes something internal and reversible. If your task speaks with a client, expect N2.
Check your understanding
What is the main difference between using AI as a tool and working with an AI agent?
What you take from this module
Explain in your own words what an AI agent is and do your first delegation with a calculated level of autonomy.
- What is an AI agent.
- The turning point: before you executed; now you manage.
- Intent-driven automation.
- The new workflow.
- People + Processes + Agents.
- Your first delegation today.
Next action: apply what you learned to your agent sheet and note what still needs review.