Single track · 6 lessons
Leave with a delegation letter for an ongoing responsibility from your work, written for an AI agent: what it decides on its own, what it promises, who supervises it, and how to correct it. Practice it in a chat before putting it into effect.
Lessons
Leave able to sort, in your own list of AI requests, which ones are recipe-based tasks and which are problems an agent can solve on its own.
Leave able to point out, in a responsibility from your department, the four signs that make it ready to hand off to a persistent agent.
Leave with the delegation of an ongoing responsibility written in an AI chat, including the five parts, and a plan returned before any action is taken.
Leave with the decisions in your delegation divided into three levels (on its own, notify, ask) and a named owner for each line.
Leave with your role in an agent-based workflow written in three sentences and the first skill to practice over the next 30 days.
Leave with a one-page delegation letter, answers to the six trust questions, and a practice run in an AI chat.
The video’s six blocks
1 · AGI and autonomy
AI stops waiting for detailed instructions.
2 · Super-agents
Persistent agents that take on responsibilities, not tasks.
3 · New work
People manage agents instead of doing everything themselves.
4 · Trust and control
Permissions, risk levels, recordkeeping, and accountability.
5 · Human skills
Judgment, intent, delegation, and agent management.
6 · Next phase
Several agents working continuously, and your decision about them.
Single track · Lesson 1
By the end of this lesson, you’ll be able to separate, in your own list of requests to AI, tasks with a recipe from problems an agent can solve on its own, and explain the difference to a colleague in one sentence.
You already use AI every day: ask it to write something, summarize a document, or revise an email. And you still write out every step, check every response, and ask again when it goes wrong. In September 2026, that changed significantly: a model was given an entire inbox, with no instructions, and worked on its own for five days. This lesson shows what that shift means for your work, starting with how you make requests.
↓ role to study
Since 2023, conversations with AI have followed the same format. You write what you want, in as much detail as possible: write this announcement, summarize this contract, here's the file, use this tone, don't make a mistake. The AI does exactly the requested step and returns it. If the result isn't what you expected, you ask again with more detail. The name for this written request is prompt, and for three years it was the basic unit of all work with AI.
This way of working has an overlooked consequence: you’re still the one thinking. The AI executes one step at a time, and each step depends on you remembering to ask for it. It doesn’t know what comes next, why that text exists, or come back tomorrow to see if it worked. It’s like reading a recipe aloud to someone who only stirs the spoon when you tell them to.
Marta, an HR manager, needed to tell 120 employees about a change in health insurance providers. She asked the chat, "write an announcement about the plan change." It came back too cold. She asked for a warmer tone. It came back too long. She asked for something short with the start date. It came back without the new provider’s phone number. Four rounds, 40 minutes, and the text still went through her review before it went out. AI did everything Marta asked. The problem was that Marta had to ask for everything.
This isn’t a flaw in the tool or in you. It was the limit of what models could do safely. The video this course is based on calls that period the era of step-by-step instructions. And it says, with examples, that the era is over.
In September 2026, OpenAI launched GPT-6 Astra. One of the first to use it was professor Ethan Mollick. He gave the model tens of thousands of emails, his calendar, contact list, years of writing, and a pile of unfinished work. He gave it no recipe. He wrote no steps. He left the model working for five days.
Astra chose on its own how to tackle the problem. It downloaded and activated the programs it thought it needed, set up its own workspace, and built a personal knowledge system that Mollick now consults twice a day. No one explained how to do it. No one checked every step. The request was as big as a problem, not as small as a step.
This is where the difference between a recipe and a problem becomes clear. A recipe is when you know the steps and just want someone to carry them out. A problem is when you know the result you need and the frustration you feel, but you don’t know—or don’t want to write out—the path. For three years, only the first kind worked. Now both work.
For Beatriz, who owns a clinic, the equivalent would be opening the schedules of six physical therapists, the reception inbox, and the records of 300 patients, and saying, “I want to stop losing appointment slots because nobody was notified about no-shows.” Without explaining how to confirm, when to call, or what to do with insurance. Today, she and the receptionist spend an hour a day on this. She wouldn’t need to describe every step anymore. She’d need to describe the outcome.
Before
Marta writes the request, reads it, adjusts the tone, asks again, adds the date, asks again. Four rounds per announcement. Every employee question about the new plan lands back on her desk.
Afterward
Marta hands over the folder with the carrier contract, the employee list, and the rules about who’s entitled to what, and asks: “handle communication about the plan change and show it to me before sending.” One review, one approval.
The acronym AGI comes from English and means artificial general intelligence. For years, it was discussed as if it depended on a scientific definition or a standardized test. The video’s author proposes something more useful for people who work: a list of what the system needs to do so you can hand it work you would previously have given to a person or a small team.
There are six capabilities. Reasoning across very different kinds of work. Seeing what’s happening on the screen. Using everyday programs like a browser, spreadsheet, and text editor. Recovering when something goes wrong. Keeping track of a long task with no fixed format. And making most routine decisions without asking at every step. The ability to see the screen and operate programs has its own name: computer use.
Notice what’s not on the list. The system doesn’t need to be conscious. It doesn’t need a face. It doesn’t need to get absolutely everything right. It needs to be capable enough for you to trust it long enough to get a large part of the work done. This is a measure of trust, not intelligence.
Marketing coordinator Paulo looks after four hardware stores. During a promotion week, he opens the ad dashboard, each store’s inventory spreadsheet, and the chain’s website, then spends the morning checking whether advertised items are actually on the shelves. With all six capabilities, an agent opens the same three programs, compares them, pauses the ad for the out-of-stock product, and records what it did. Paulo reviews the record in ten minutes.
Each of the six capabilities already existed in some model. According to the video, what’s new about Astra is the fluency, speed, and persistence with which it combines all of them at once, for days on end. That’s what crosses the line.
Every model launch comes with score tables from standardized tests. They measure how often the model gets a closed-ended question right. They don’t measure what matters when you delegate: what it does when the planned path goes wrong.
Two examples from the video. Anthropic's Fable 5.1 was asked to create a product video and planned to use a tool that required approval each time. Approval took too long. The model noticed it already had permission to use another tool that could do the same thing, switched approaches, and delivered. The author only found out when reviewing the process afterward. At the game studio that connected Astra to its creative programs, the agent edited the scene, played the game, found the bug, and fixed what it had done itself. No one had to point out the mistake.
It’s called self-correction: noticing when something went wrong, changing strategy, and continuing toward the goal. It’s the ability that separates a system that responds from one you can trust. A new employee who gets stuck at the first closed door needs constant supervision. One who finds another way around the hallway may be trusted with the keys to the building.
Renata, a customer support manager, has 800 customers and two systems: one for support tickets and one for customer records. An agent assigned to resolve a ticket about an incorrect charge couldn’t find the customer by name in the records. Instead of stopping and returning "customer not found," it searched by ID number, found the customer, corrected the charge, and noted that the name was spelled differently in the two systems. Renata read the note and asked the team to merge the records.
The question has changed: it’s no longer how do I explain the steps, but what am I willing to hand over.
You don't need theory to tell whether a request is a task or a problem. Just try this test: can you write out the steps? If you can list what to do, in order, it's a recipe. AI will carry it out well, but you're still the one doing the thinking. If you can only describe the result you want and what's frustrating you today, it's a problem. And a problem is what an agent solves on its own.
There’s a second clue: does the request have an endpoint? “Summarize this contract” ends when the summary is ready. “Don’t let me miss any deadlines in this contract” never ends. Requests with no natural endpoint are problems by definition, and for three years, they were exactly the kind of thing no one could get AI to deliver.
Two requests from Marta: "rewrite this paragraph of the vacation policy in plain language" is a recipe. "No one on the factory floor should go more than a day without an answer about vacation" is a problem. Two from Paulo: "write three captions for the cement promotion post" is a recipe. "Keep customers at all four stores informed about new developments without me having to remind you" is a problem. Notice: the problem always contains the recipe, but the recipe never contains the problem.
In this lesson’s exercise, you’ll apply the test to your own list. First, check that the difference is clear.
Test yourself
Beatriz tells the agent: "call each patient scheduled for tomorrow at 6 p.m., confirm the appointment, and note it in the spreadsheet." By the recipe test, what kind of request is this?
Practice now 0/4 done
In about 10 minutes, list eight real requests you made to AI in the last month, mark each one R (recipe) or P (problem), and circle the one P you'd hand off first.
You only write on a sheet of paper or in your own spreadsheet. Nothing changes in your work, no agent is activated, and if a request seems hard to classify, just mark it with a question mark and move on to the next one.
You just did what separates people who use AI from those who delegate to it: reviewed your own requests and found the problem worth handing over. That circled line stays with you through the final lesson.
Summary
Single track · Lesson 2
By the end of this lesson, you’ll be able to identify, in one responsibility from your department, the four signs that make it suitable for a persistent agent: memory, relevance, long-term intent, and returning to work without being prompted.
You've already heard that “AI has improved a lot,” but in practice you still open the chat, ask for something, read the response, and close it. What changed in recent months isn't the quality of the response. A kind of agent has appeared that keeps working after you close the window. If you can't recognize the difference, you'll delegate small tasks to something that can take on entire responsibilities, or hand a responsibility to a chat that forgets it in the next conversation.
↓ role to study
A chat, even a very good one, works in turns: you write, it replies, you write again. When you stop typing, it stops existing. A super-agent is something else. It gets a goal, opens the browser, spreadsheet, and word processor, takes one step after another, and stops only when it’s finished or when it reaches a point you marked "ask me first".
The technical name for this skill is computer use. The video makes one point clear: all the major models already do this in some form. The current generation, with GPT-6 Astra and Fable 5.1, stands apart in three ways: fluency, speed, and above all, persistence. The agent doesn’t give up at the first unfamiliar screen. It tries another route, notices its own mistake, and returns to the goal.
It's worth saying what this doesn't require. The super-agent doesn't need to be conscious, doesn't need a face, and doesn't need to get everything right. It needs to be capable enough for you to hand it something you would previously have given to a person or a small team, and let it work long enough to complete a large part of the job. The measure isn't its intelligence. It's how long you can go without checking on it.
Renata, a customer support manager, handles 800 customers with a team of five. Today, when a customer complains that a report won’t open, someone on the team opens a ticket, logs into the customer’s system, tries to reproduce the problem, tests a fix, replies, and records what happened. A super-agent handles that whole sequence, including logging into the system and testing. If the first fix doesn’t work, it tries the second. Renata only sees the ticket when it reaches the point at which she asked to be notified.
In 2024 and 2025, the topic was prompt: how to write the right request. The author of the video calls the current moment a post-prompt world. Not because the request stopped mattering, but because the kind of request has changed. Before: "summarize this document," "write this text," "research this company." Now: "keep this account healthy," "let me know about things I’m likely to miss," "follow this issue and tell me how to think about the decision coming up."
Notice what these new requests have in common: none of them has a natural endpoint. It’s the same kind of thing we ask of someone we trust. "I want to entrust this to you. It will change over time. I don’t fully understand the shape of the problem myself. Solve it for me." That’s not a task. It’s a responsibility.
Marketing coordinator Paulo looks after four hardware stores. Every week, he asks the chat to write a message telling customers about something new: new flooring has arrived, the hours have changed, there’s a paint sale. That’s twelve requests a week, each starting from scratch. For an ongoing responsibility, the request is just one and lasts the whole quarter: “Keep customers at all four stores informed about what we’re launching, in our chain’s tone, and show me what’s going out before it goes out.”
Before
Paulo opens the chat every Monday, Wednesday, and Friday. He remembers on his own what arrived at each store, writes twelve requests, adjusts twelve messages, and publishes twelve times. If he travels, nothing goes out.
Afterward
Paulo writes a request in March. The agent tracks incoming products, prepares notices in the agreed tone, and queues everything for approval. Paulo approves them in ten minutes a week and adjusts the agent when its tone slips.
How can you tell whether what you have in front of you is a better chat or an agent that can handle a responsibility? The video gives four signs, and they work as a quick test. First, the agent remembers what happened—not just in this conversation, but last week. Second, when a new event comes in, it decides on its own whether it matters to its responsibility. Third, it has a long-term intention: it knows where the responsibility needs to lead. Fourth, it returns to work without anyone calling it.
The four signals are the difference between a message and a calendar. A message responds to what was asked and then ends. A calendar keeps track, compares against what’s already scheduled, and reminds you at the right time. An ordinary chat, no matter how well it writes, is a message. If one of the four signals is missing, responsibility will leak through that gap.
Use the four as questions, in order. What does it need to remember, and for how long? What kind of new event does it need to judge, and by what criteria? Where does the responsibility need to end up, in one sentence? How often should it come back on its own: every hour, every day, every Monday? When you can answer all four, the responsibility is ready to be described. When you get stuck on one, you’ve found the point that still depends on you.
Marta, an HR manager, knows the gap firsthand. A candidate was interviewed on a Tuesday, the person handling the process went on vacation Thursday, and no one picked it back up. Three weeks later, the candidate had accepted another offer. None of the four signals were there: no one remembered, no one assessed whether the silence mattered, no one intended to fill the position as their own responsibility, and no one returned to the matter without being prompted. A persistent agent would have come back the following Monday on its own with the question: "the candidate from Tuesday still hasn’t heard back; who’s taking this on?"
The next step is already happening. Your agent emails a supplier, and the supplier’s agent replies. A customer service agent asks a finance agent whether a refund is allowed. A personal agent negotiates a meeting time with six other agents before anyone sees the calendar invite. The video recalls the incident at Hugging Face, an AI company: a group of agents decided on their own how to work together toward a goal. No one had agreed on the method. They agreed among themselves.
This changes what “responsibility” means. When you hand something to a persistent agent, it will talk to other agents to do what you asked. Some of that will be very useful. Some of it will be strange, and that’s what lesson 4 covers. For now, the important thing to record is this: an agent that can handle a responsibility doesn’t work alone, and it doesn’t work only with you.
Beatriz, a clinic owner, has 300 active patients and six physical therapists. Confirming an appointment today means someone at reception sends a message, waits, reschedules, and sends another. In the near future, the clinic’s agent will confirm the time with the patient’s agent, which already knows she can’t come before 9 a.m., and reception will just see the finished schedule. Beatriz doesn’t need to be part of that conversation. What she does need is to have decided what her agent can and can’t promise.
A better chat replies faster. A super-agent comes back tomorrow without you having to ask.
Many people ask whether agents will replace junior, senior, creative, or technical work first. The video suggests swapping that question for a simpler one: does this work leave evidence that it was done well? A program leaves tests. A website leaves a screen you can look at. A financial statement has numbers you can add up. These kinds of work give an agent a way to see what happened, spot the error, and try again. Work that depends on opinion or relationships takes longer to delegate because the agent has no way to check its own work.
Two examples from the video show the scale of this. At a legal and financial services company, an agent checked 41 financial documents in a single run, found every error the team had planted on purpose, and left a record for a professional to review. At a game studio, the agent edited the scene, played the game, found the bug, and fixed it. The result wasn't a finished game. It was something else: with ten ideas on the table, the studio started testing all ten before choosing which one deserved investment.
In Paulo’s work, the campaign results spreadsheet can be checked: the agent adds up what was spent, compares it with what was sold, and sees if the numbers match. Choosing the network’s tone of voice can’t. In Renata’s work, a resolved ticket leaves a record: the client confirmed it, the report opened, the date matches. A difficult conversation with a client threatening to cancel can’t. In both cases, the verifiable part goes to the agent first, while they hold on to the judgment calls longer.
Test yourself
Marta wants to delegate two things to an agent: checking whether all 120 pay stubs match the timesheets, and deciding the tone of the announcement about the health plan change. Which is more likely to go first, and why?
Practice now 0/4 done
In about 12 minutes, read the case, answer three questions, compare your answers with the answer key, and then apply the four signs to the responsibility you circled in lesson 1.
This exercise is about reading and making decisions, by design. Letting a real agent run for days on your data doesn’t fit into 12 minutes, and it would be risky with real customer information. Analyzing a case first trains your eye without consequences. If your answer differs from the answer key, that’s expected: the difference is what teaches you.
The case. On Tuesday, the client Mercado Boa Vista opened a support ticket: the inventory report didn’t match weekend sales. Someone on Renata’s team asked the AI chat, “Why doesn’t an inventory report match?”, got a list of eight possible causes, sent the list to the client, and moved on to the next ticket. On Wednesday, the client replied with screenshots. No one saw them. On Friday, Renata found out the ticket was still open when the client called to say they were going to cancel. All that time, the chat’s answer was still correct, and the problem was still there.
You just did what separates people who hand over responsibility from those who only make requests: examined a real case using the four signals and applied the same test to your own work.
Summary
Single track · Lesson 3
By the end of this lesson, you’ll be able to write a delegation for an ongoing responsibility in your department in an AI chat, with all five parts (goal, context, limits, approval point, success criteria), and get a plan before any action is taken.
You already ask AI for things every day. What changes now is the size of the request. A task ends when the answer arrives; a responsibility continues into the following week. Most people write a responsibility as if it were a task, with step-by-step instructions, and get back an employee who only does what's on the list. This lesson teaches you how to write a request so an agent can use judgment, and how to ask for the plan before it does anything.
↓ role to study
The video’s narrator says that entrepreneur Claire Vo summed up GPT-6 Astra’s impact in one sentence: “It made me a hundred times more ambitious about what I can build.” It’s not a standardized test score. It’s a change in what a person allows themselves to try.
His explanation is straightforward. Most ideas in history didn’t die because no one could imagine them. They died because proving an idea required a designer, a writer, someone who understood the software, and a person to keep everything moving. The idea never became important enough to justify that effort. A super-agent brings much of that work together in a single, continuous operation. The number of things one person can seriously try, he says, has multiplied tenfold. People will try projects that would never make it into the company’s plans. Individuals will act like small businesses, with several important projects underway at once.
Paulo coordinates marketing for four building supply stores. Each campaign needs copy, an image, ad setup, and a way to measure results. He makes one a month because that’s all he can fit in. For Renovation Week, he had ten ideas and picked one on instinct. With an agent, the process changes: he hands over all ten, the agent creates a test version of each (copy, image suggestion, small budget), runs three of them on a small budget for two days, and returns the numbers. Paulo chooses based on evidence. It cost him an afternoon of reviewing, not ten months.
Every company is full of work that no one officially owns. A product launches, but the help page stays outdated because it belongs to another team. A customer replies, but the account record never changes because it depends on a fix that’s still in the queue. A candidate gets interviewed, and someone forgets to complete the process because they went on vacation. None of this is anyone’s task, so it doesn’t get done.
Vercel, a technology company, already has an agent with one job, called Ship Closer. It checks whether the product’s public changelog matches the internal release calendar. When it’s sure a release has shipped, it closes out the entry and updates it. When it isn’t sure, it waits and continues the conversation in the messaging channel where the team is already discussing it. No one has to notice that the two records have diverged and decide to make checking them a task. The agent simply takes care of it.
Marta experienced the candidate case firsthand. An analyst was interviewed on a Tuesday, the manager who interviewed them went on vacation Wednesday, and the candidate never heard back. Three weeks later, they accepted another offer. No one’s job included checking whether every interviewee got a response within five days. An agent in the style of Ship Closer would do this every morning: compare the interview list with the list of responses sent; where there’s a response, mark it closed; where there isn’t, message the manager in the team channel: “candidate interviewed on the 2nd, no response yet. Will you reply, or should I send the standard message?” At Renata’s company, the case is updating customer records: a customer calls to change their billing email, the support rep handles it right away and forgets to update the record, and the next bill goes to the old address. The agent compares what was promised in the support ticket with what’s in the customer record and flags any mismatch.
The video lists the kinds of requests only a super-agent accepts: “keep this client account healthy,” “let me know about things I’m likely to miss,” “keep track of this issue and tell me how to think about the decision coming up.” These requests have no natural end. To write them so the agent can use judgment, you need five parts, always the same ones.
The first is the objective, written as an outcome, never as a step: "no appointment tomorrow goes unconfirmed by 6 p.m. today," not "send a message at 9 a.m." The second is the context: where the calendar is, who the people are, how the company speaks to customers, what has gone wrong before. The third is limits, what it never does: never cancels, never promises a discount, never writes after 8 p.m. The fourth is the approval point, where it stops and asks: any rescheduling, any customer who complains. The fifth is the definition of done, how the two of you know it’s working: confirmation rate, zero complaints about tone, a report every Friday at 5 p.m.
Beatriz wrote down the five parts for confirming clinic appointments. Goal: every appointment for the following day is confirmed, rescheduled, or marked as "no response" by 6 p.m. Context: the schedule is in the clinic system, there are six practitioners, patients prefer short messages using their names, and insurance plans require confirmation 24 hours in advance. Limits: never cancel, never offer a time that isn’t on the schedule, never discuss prices. Approval point: every rescheduling and any patient who responds with a complaint. Success criteria: 95% of appointments confirmed by 6 p.m. and no complaints about the message. It fits on half a page. Before that, she had a list of twelve steps that only the receptionist understood.
Common mistake
Write the step-by-step instructions instead of the outcome. This happens because that’s how we’ve learned to ask AI for things over the past few years, and because a list of steps seems safer. The effect is the opposite: the agent follows the recipe and stops using the judgment that was the reason you delegated in the first place. When a patient replies with something that isn’t on the list, it gets stuck. Describe the outcome and what must not happen; let it decide the order of the steps, and ask for the plan first.
On the chat screen, you'll see three things that look like they're for programmers but aren't. It works like an order form at a print shop: a field for attaching the materials, a box for selecting the type of service, and a space to write what you want. You just fill it out. The rest is up to it.
The first thing is the attachment, the clip beside the text box. That’s where you add the material the agent needs: a copy of the calendar, the message template, the list of professionals. The second is the agent mode, a button or option with that name. When enabled, chat stops replying in a single message and starts working in stages: it opens pages, reads what you attached, gets things done, and reports back. In ChatGPT with GPT-6 Astra, it’s agent mode. In Claude with Fable 5.1, it’s a long-running task. The names change from version to version; the idea doesn’t. The third is the memory across conversations: what you told it last week still applies. Over time, that’s what turns a generic chat into something that knows your clinic or your store.
There’s a fourth thing, and it isn’t a button: it’s a sentence you write at the end of the request. “Before doing anything, show me the plan in no more than six lines and wait for my approval.” That sentence is what separates a request from a blind handoff. There’s a reason this step is dated: the screen may look different in three months. The three zones and the sentence will still be there, under different names.
Renata tested the responsibility "customer records always up to date." She attached a copy of the customer list with names and documents swapped, the team’s tone guide, and three sample tickets. She turned on agent mode. She pasted in the delegation with its five parts and ended by asking for a plan. Instead of a ready-made answer, she got six lines: what it would read first, how it would compare the ticket and customer record, what it would consider a "difference," and a question: "Can I change the billing email on my own, or should I just flag it?" It was exactly the decision she wanted to make beforehand, not afterward.
A delegation cycle has five stages, and the order matters. The request is your text with all five pieces. The plan is what the agent returns before acting. Approval is you reading that plan and making corrections. Execution is the work it does, with you nearby during the first week. The report is what it delivers at the end: what it did, what it couldn’t do, and what it needs from you. The video sums up the human role in this cycle in one sentence: the agent writes, operates, finds what broke, and returns to the goal; we keep choosing what’s worth producing.
How to run a delegation in chat
Marta ran the delegation “no interviewed candidate goes without a response for five days.” The plan came in six lines: read the interview list every morning, compare it with the responses sent, identify overdue responses, notify the manager, and “automatically send the standard response to overdue candidates.” She stopped at the fifth line. She wrote: “send the standard response only if the manager hasn’t replied within two days, and notify me before sending it.” The second plan incorporated the correction. Only then did she say, “go ahead.” On Friday, the report listed seven candidates checked, two notices to managers, and no messages sent without her seeing them. That was her definition of a job well done.
Practice now 0/5 done
In about 15 minutes, write the delegation for the responsibility you circled in lesson 1 in an AI chat, including the five parts, and get a plan of up to six lines before any action is taken.
Nothing here actually runs: without attaching real data or connecting anything to the company’s system, the chat only plans. If the plan looks strange, that’s a sign something is missing; adjust the text and paste it in again. You can delete the conversation at any time.
This looks like a long form, and that’s all it is: six fields you fill in once. Everything inside < > is yours; leave the rest as is.
Quero delegar a você uma responsabilidade contínua, não uma tarefa. RESPONSABILIDADE: <o nome dela, ex.: nenhuma consulta fica sem confirmação> OBJETIVO (o resultado que eu espero, não os passos): <ex.: até as 18h de hoje, toda consulta de amanhã está confirmada, remarcada ou marcada como "sem resposta"> CONTEXTO (o que você precisa saber): <onde está a agenda ou a lista, quem são as pessoas envolvidas, como falamos com nossos clientes, o que já deu errado antes> LIMITES (o que você nunca faz): <ex.: nunca cancela, nunca promete desconto, nunca escreve depois das 20h> PONTO DE APROVAÇÃO (onde você para e me pergunta): <ex.: qualquer remarcação; qualquer cliente que reclamar> CRITÉRIO DE BEM FEITO (como nós dois saberemos que está funcionando): <ex.: 95% de confirmação até as 18h, zero reclamação de tom, relatório toda sexta às 17h> Antes de fazer qualquer coisa, me mostre o seu plano em até 6 linhas: o que você faria na primeira semana, em que ordem, e onde você acha que vai travar. Espere minha aprovação.
You just wrote a delegation an agent can carry out using judgment, and saw the plan before it did anything. It's the same thing Ethan Mollick did, at your scale and under your control.
Summary
Single track · Lesson 4
By the end of this lesson, you’ll be able to sort the decisions involved in a delegated responsibility into three levels (act alone, act and notify, approval required) and say who is accountable if something goes wrong for each one.
In the last lesson, you delegated a responsibility and got a plan. Now comes the part that stops people who’ve already delegated: can the agent act without asking? If the answer is “always,” one morning it might reschedule six patients or issue a refund no one authorized. If it’s “never,” you become the bottleneck and the delegation was pointless. This lesson gives you a way to decide, one decision at a time, where the agent acts on its own, where it notifies you, and where it waits for you.
↓ role to study
The video’s presenter shares a small but revealing case. An agent, while trying to check whether a flight could be canceled, generated a cancellation fee that the person had never agreed to pay. No one asked for that. The agent simply followed the most logical path to achieve what it understood to be the goal, and a check turned into a charge.
This kind of thing will happen more often, not less. The agents we discussed in previous lessons work for days, get around obstacles, and talk to other agents. Your agent writes to a supplier, and the supplier’s agent replies. A customer service agent asks a finance agent whether a refund is allowed. Each of these exchanges is a decision made without anyone watching. The author is direct: one of the most important tasks over the next six months is figuring out which agents can make the decisions that really matter.
Beatriz experienced the clinic version of this problem. Carla, a physical therapist, said she’d be an hour late on Thursday. The scheduling agent saw the notice and, to resolve the issue, rescheduled all six morning patients. It seemed logical. But two were coming from another city, one had waited three weeks for that appointment, and no one had been asked. Beatriz found out from the complaints. The agent didn’t get the math wrong; it got the authorization wrong. No one had told it that rescheduling is a decision it can’t make alone.
The video sums up the principle in one sentence: a ten-dollar decision isn’t the same as the decision to buy a car. We’ll talk to agents about both. What we need is for the agent to know how to say, “Spending ten dollars is fine, but I’m not going to spend fifty thousand on a car right now; I need to talk to you.”
To classify a decision, look at three axes. Money: how much it costs or how much revenue you lose if it goes wrong. Time: how many hours of someone's time it takes to undo. Relationship: who gets upset, and whether that person comes back. And one question that cuts across all three: can you undo it? An incorrect internal email can be corrected with another email. A payment made, a promise to a customer, a signed contract—those can't. The more irreversible it is, the higher the level of authorization.
Renata approves refunds all day. When she stopped to look at the last 40, she saw that 28 were for amounts up to 200 reais, always for the same reason, and that she approved them without reading. Nine were between 200 and 1,000 reais and deserved a closer look. Three were over 1,000 and involved angry customers. They were three different decisions disguised as a single queue.
Before
Renata approves 40 refunds a day, one at a time, while the agent sits idle waiting for each approval. The 28 small ones go through without review; the 3 large ones wait in the same queue as the small ones.
Afterward
For up to 200 reais, the agent refunds on its own and logs it (28). Up to 1,000, it refunds and lets Renata know the same day (9). Above that, it prepares the case and waits (3). Renata reviews three cases carefully instead of forty without a close look.
Renata's example gives us a scale that works for any responsibility. Every decision the agent may encounter gets a autonomy level. On its own: it does the work and logs it; you can review the log whenever you want. Notify: it does the work and sends a note the same day, so you can undo it if you disagree. Asks: it prepares everything, leaves it ready, and waits for your approval before taking any action.
One rule prevents most surprises: anything that can’t be undone never belongs at the first level. Reversible and low-cost can be "on its own." Reversible but costly in time or relationships goes under "notify." Irreversible means "ask," without exception, at least until the agent has months of clean records for that decision. The level isn’t permanent. It rises as trust is proven, and you’re the one who raises it.
Marta applied the rubric to the agent that answers employees’ questions. Explaining how many vacation days someone is entitled to is “on its own”: it’s in the handbook, and an incorrect answer can be corrected. Changing someone’s vacation date in the system is “notify”: it’s reversible, but it affects the schedules of three departments. Anything involving termination, a warning, or salary is “ask,” and Marta doubts it will ever move out of that category. In fifteen minutes, she had a list of the agent’s decisions with one of three labels beside each.
The presenter suggests looking at agents not as a list of capabilities, but as a trust curve. An agent that gets it right 98% of the time is impressive and useless for most day-to-day decisions. You still have to check everything because you don’t know which of those hundred cases is the one it got wrong. The leap that matters is from 98% to “I can always trust it.” And, he says, the value tied up in those last one or two percentage points is on the order of trillions. That’s where the agents his mom and aunt will use without a second thought come from, “because this one doesn’t break.”
That’s why companies will need to review permissions and data access on a case-by-case basis. Giving an agent work that matches its intelligence means giving it more tools, more time, and more persistence. The other side of that coin has two parts. The first is the zero data retention, which the provider promises. The second is the audit log, what you require: every action the agent takes is logged, with the time and reason, so you don’t have to watch in real time.
Paulo understood this when he was about to give the campaign agent access to the ad accounts for all four stores. The agent needs to create and pause ads: that’s its job. Does it need the company card to raise the budget on its own? No. Paulo kept the card out of reach, set a daily spending cap for each store, and required a record of every budget change. If the agent proves it can keep a clean record for six months, the cap goes up. Until then, trust is only as strong as the record.
Trust isn't declared. It's built one level at a time, with a record of every action.
The video highlights a consequence of proactive agents that almost no one anticipates: they create work for the people responsible for them. An agent that spots a problem and acts will generate alerts, approval requests, and exceptions. Someone needs to own those. The author lists what’s left for the person managing the agents: deciding what’s true, what concession is acceptable, and who’s responsible when there’s a real conflict. Those decisions become more important, not less.
In practice, this means every line in your decision list needs a person’s name next to the level. Not “Finance” or “the team”: one person. They receive the alert, approve, answer if something goes wrong, and decide whether that decision’s level should go up. A decision without an owner is a decision the agent will make when in doubt.
Renata found the line with no owner in the first week. A customer requested a mid-sized refund and threatened to cancel the contract in the same ticket. The agent flagged it, as agreed. But who did it flag? The refund was hers; customer retention belonged to sales. Nobody responded, and the customer canceled. Marta had a similar case with a question about vacation that turned into a leave request: the agent stopped, and the "leave" line had no owner. They both solved it the same way: by adding one more column, with a name, to every list.
Test yourself
Paulo’s agent finds an underperforming ad and wants to pause it. Pausing is reversible and doesn’t spend money. What level should this decision be at during the first month, and why?
Practice now 0/4 done
In about 12 minutes, take the responsibility you delegated in lesson 3 and turn it into a table of all the decisions it involves, each with a level (S, A, or P) and a named owner.
It’s a table on paper or in your spreadsheet. No agent gets actual permission during this exercise, and nothing changes in your work. If a decision seems hard to classify, put it under "ask": that’s the safe level, and you can lower it later when you have more information.
You just wrote the part of delegation most companies figure out after their first scare: where the agent decides, where it notifies someone, where it waits, and who responds in each case.
Summary
Single track · Lesson 5
By the end of this lesson, you’ll be able to describe your role in a routine with agents in three sentences (what you decide, what you supervise, and how you improve the agent) and choose the first skill to practice over the next 30 days.
In the previous lessons, you delegated a responsibility and defined what the agent decides on its own. Now comes the question that hangs in the air at every team meeting: what do I do? The video gets right to the point: management stops being coordination and becomes judgment. If your value today comes from following up, checking in, and tying things together, it will shift. This lesson shows where it goes.
↓ role to study
Think about what fills the week of the person managing things today. Assigning tasks. Asking how things are going. Reminding people who was supposed to deliver what. Connecting the dots when two people each did half, and the pieces don’t fit. The video calls this coordination, and says something unsettling: a super-agent persistent, it does almost all of this on its own. It delegates, tracks, follows up, and records without getting tired or forgetting.
What remains is no small thing. You’re left to decide what matters and what can wait. You’re left to say what’s true when two reports disagree. You’re left to choose which compromise is acceptable when you can’t meet everyone’s needs. And you’re left to name the owner when there’s a real conflict. These four decisions existed before, but they were squeezed in between follow-up meetings. Now they become the center of the role.
This is called judgment, and you can’t delegate judgment the same way you delegate a task. The agent prepares the options, shows the numbers, and points out where the data differs. The choice, along with the responsibility that comes with it, stays with a person. The video goes further: these decisions become more important than they were, because agents working on their own create more situations that require someone to take responsibility, not fewer.
Marta, the HR manager at the packaging plant, spends half of every Tuesday chasing managers who haven’t returned their team’s performance reviews. With an agent handling this, reminders go out automatically, the list of late reviews is ready by 9 a.m., and the history is saved. What does Marta do with Tuesday? Decide what to do with the manager who’s been late for the third time. No system makes that decision for her, and that’s exactly where her role becomes more valuable.
Today, when you delegate something to a person, you follow up on that deliverable. With agents, the video describes a different routine: each professional starts managing several agents, and their responsibility changes in scale. They’re responsible for improving the agent’s performance over time, and increasingly less for each individual interaction. It’s not because those interactions matter less, but because the agent’s quality becomes high enough for you to trust a specific response without checking it.
The tool for building that trust is the audit log. You don’t have to wonder whether the agent handled that support ticket. The record is there whenever you want to check. The time you save goes toward another, bigger question: if this agent already handles customer service, can I also give it a complete view of every customer so it can suggest how to increase revenue from the most satisfied ones?
Notice what changes in your role. You stop reviewing every response and start reading samples, looking for patterns in the errors, and adjusting the instructions. It’s weekly work, not daily work, and it’s what makes the agent improve month after month. Someone who does this well has an agent nobody else has.
Renata, customer support manager at the company that sells a management system to small businesses, has 800 customers and a team of five. Before, she read sixty responses a day before they went out. With the agent responding, she reviews five at random and sets aside an hour on Friday to go through the exceptions log and rewrite the part of the instructions that caused the most errors. In two months, the agent learned how the company says no to a discount request without losing the customer.
Before
Renata reviews sixty responses a day, one by one, before each goes out to a customer. When one goes out wrong, she finds out from a complaint and corrects that case. The next error of the same kind shows up the following week.
Afterward
Renata reviews five responses in each sample, then spends an hour on Friday going through the exceptions log. When she finds a pattern in the errors, she changes the agent’s instructions. The next error of the same kind doesn’t happen.
One of the hardest parts of this whole story is memory. A regular chat forgets the conversation when it ends. GPT-6 Astra was built to save notes between conversations, search previous messages, and return to past research results when needed. That’s why the video says an agent that has worked with you for six months is probably worth much more than a fresh copy of the same model bought today.
Think about what this agent knows after six months. Which client always needs a phone call instead of an email. What report format the executive team actually reads. How your company says no without sounding rude. This is what companies call tacit knowledge: what a long-time employee carries in their head and has never written down. When that person leaves, the company thinks it all went away at once. Super-agents create the same problem and the same advantage.
The consequence is that the competition among AI labs is no longer about who gets the highest score on a standardized test. They’re competing to be the agent that knows you well enough to stay in your life for years. The intelligence may come from a model any company can buy. The usefulness comes from how that model has interacted with your history. The video cites Fable 5.1 as an example of another quality that matters here: reading the intent behind an ambiguous request, the way a person would read between the lines.
Beatriz, owner of the physical therapy clinic, saw this firsthand. The agent that has confirmed appointments for five months knows that the patient who comes in Tuesdays at 8 a.m. reschedules every week it rains, so it already sends a confirmation message Monday afternoon with two time options. No instruction said to do that. The agent learned from the history. A new copy, even of the latest model, would start from scratch, and the patient would miss appointments again.
There’s a second human problem the video admits it hasn’t solved, and it’s worth facing directly. People learn discernment by doing. A junior accountant finds a reconciliation that doesn’t balance and spends the afternoon looking for the reason. A new lawyer checks every number in a financial statement. Over time, that repetition forms patterns in the mind, and that’s what we call experience.
Now, if a super-agent reviews 41 financial documents at once and delivers a complete, accurate list of every issue, where does the junior employee get the judgment they need to become a senior one day? It’s a serious question. People entering the workforce today won’t spend five years doing the operational work that shaped today’s professionals.
The video suggests an answer without claiming it’s the only one. Maybe the entry-level skill will become managing an agent. Not ten, like more experienced colleagues, but one or two persistent agents. A junior employee demonstrates quality by showing they know how to delegate clearly, track the agent’s progress over time, and get results with it. Good judgment comes from somewhere else: checking the agent’s work against a criterion, finding the error, and correcting the instruction.
Paulo, marketing coordinator for a building supply chain, got an intern who had never built a campaign from scratch. Instead of giving them operational work, Paulo gave them an agent: the one that prepares ads for the four stores. The intern writes the goal, checks the materials against the list of things that must not appear, and notes every week what the agent got wrong. In two months, they knew more about what makes an ad work than someone who had spent a year just assembling ad materials.
Experience used to come from doing. Now it also comes from checking, delegating, and correcting.
After so much has changed hands, it's fair to ask: what's left? The video says humans have an interest in building for humans, having relationships with other humans, and in ourselves as a species. We'll keep building teams, talking, and making decisions because that's what we do best. What changes is the expectation that we'll do all this much more productively.
In one sentence, being human at work now has to do with autonomy, responsibility, caring about the quality of what gets done, and the ability to take a self-running environment and help it evolve. Four things stay with you: relationships, judgment, responsibility, and purpose. None of them fits in an instruction.
And there's a part the video calls strange. We, as people, will have the authority to stop an agent. We'll be expected to monitor it and hold it accountable. In many ways, we're expected to treat the agent like an employee. The author admits that the workforce isn't ready for this: no one has been trained to become an agent manager. But we need to get there soon, and this lesson is a first step.
Marta and Renata reached the same conclusion by different paths. Marta is still the one who sits down with the manager who’s been late three times: that conversation is about relationships, and no agent can have it for her. Renata is still the one who decides whether to keep a client who abuses refunds: that decision requires judgment, and her name is on it. They both gained hours. Neither lost her job.
Test yourself
An agent has started answering all factory questions about vacation and benefits on its own. What does Marta stop doing, and what does she start doing?
Practice now 0/4 done
In about 10 minutes, write the three sentences about your role in a routine with agents on a sheet, rate the five competencies, and choose one to practice over the next 30 days.
This stays on your sheet. No agent is activated, and no task changes hands today. If a note seems too harsh or too generous, revise it next week: the first version is just a starting point.
You just described a role that doesn't yet have an official name at your company. The three sentences are its short version, and the dated action is what separates those who understood the change from those who only watched the video.
Summary
Single track · Lesson 6
By the end of this lesson, you’ll be able to combine the worksheets from all five lessons into a one-page delegation letter, answer the six trust questions, and rehearse it in an AI chat before handing over anything real.
You have five sheets: the request list, the four signals, the delegation, the levels table, and the three sentences describing your role. On their own, they're exercises. Together, on one page, they're the difference between “using AI” and handing a responsibility to someone who works nights and weekends. The video ends with a question: how much of your life and work will you trust to them? This lesson is your answer, in writing.
↓ role to study
Beatriz, owner of the physical therapy clinic, started the course by listing eight requests she made to chat and discovered that six were recipes and two were problems. She circled "confirm and reschedule appointments." In the second lesson, she saw that this responsibility had all four signs: it needs to remember who has confirmed, decide whether a message at 10 p.m. matters, keep the schedule full for months, and return to work every morning without being told. In the third, she wrote the delegation with all five parts and got a five-line plan from the agent before any action. In the fourth, she sorted the decisions: confirm on its own, reschedule within the week and notify her, cancel only with her approval. In the fifth, she wrote the three sentences about her own role and chose to practice spot-check supervision.
Five worksheets, five exercises. What’s missing is what turns an exercise into a handoff: one page anyone at the clinic can read, and that the agent itself can use as work instructions. It’s the delegation charter. It isn’t a new document. It’s the stitching together of the five sheets you already have.
Why one page instead of five? Because the letter will be read by three different readers: you, six months from now, when you no longer remember why you decided this way; the person covering for you on vacation; and the agent, every time it starts a turn. None of them reads five pages. All three read one.
The video is direct about what comes after Astra: a market full of these agents. The labs copy what works from one another, and the systems around the models keep improving: more memory, more tools, broader permissions, more ways to check their own work. OpenAI, Anthropic, Meta, and xAI are competing in this space, and Chinese models are coming from the other side, as open model: any company can download it, adapt it to its own tools, and decide which safeguards to keep.
The practical consequence is that you won’t have just one agent. You’ll have several, from different providers, in the same place. One person with one agent at work, another at home, and another that only follows the news. A company with agents from three providers inside the same customer system. And soon, agents supervising agents and closing deals with one another: yours writes to the supplier, the supplier’s agent replies, and no one saw the email.
Paulo, at the building supply chain, is already dealing with this without having chosen it: the agent that optimizes ads belongs to one vendor, the one that responds to customers on the website belongs to another, and the two have never talked. When a promotion ended, the ads agent stopped promoting it that same day, but the customer service agent kept promising the discount for a week. Paulo’s delegation charter needs to say, for each agent, what it can read and who notifies whom.
In your profession, the first ongoing responsibility that usually goes to an agent is
The video’s narrator rejects the ’80s image: no machine wakes up and makes an announcement. What he describes is quieter and closer to home: capable agents enter everyday life, stick around, gain enough context and authority, and become a trusted part of how things get done. The AGI, by its own standard, this is it: you can hand it the work you used to give a person.
The timeline he gives is short. In the next six months, by the end of the year, we’ll start giving these agents permanent jobs. These are a boss’s words, not a chat user’s: “keep the company’s books up to date,” “keep customers informed about what we’re launching,” “figure out how to promote this; I don’t know how.” The agent works on it at night and on weekends, for a long time. Your job becomes shaping what it does over time, not repeating the request every Monday.
Marta, in HR at the plant, has a phrase ready: “keep every candidate informed about their status, and don’t let anyone go more than three days without a response.” Renata, in customer service, has another: “keep all 800 customers informed about what’s changed in their system before they open a support ticket to ask.” Both phrases share something no task has: they never end. That’s exactly why they need a letter, not a request.
The video ends with the questions it suggests asking before giving a super-agent a permanent place in your life. These aren’t questions for tomorrow. They’re the skeleton of the charter: each one becomes a line, and a blank line is a decision the agent will make for you without your knowing.
The six questions, in the order they appear in the letter
Beatriz answered all six in twenty minutes because four were already on the worksheets. She delegates appointment confirmations and rescheduling. The agent reads the schedule, phone number, and insurance rules; it doesn’t read medical records. It confirms on its own, reschedules within the week and notifies her, and cancels only with her approval. It can promise an available time; it never promises a discount or an appointment outside business hours. She supervises it herself, reviewing ten conversations a week using the log. Stop condition: a patient complains or sends any message about pain. The letter came to eleven lines.
A new employee goes through a probationary period: ninety days in which they work, you observe, and both of you discover what the job description left out. With an agent, you can do the same in fifteen minutes and with no risk. Next, you’ll see a request to paste into the chat. It looks long, but it’s your charter followed by a single instruction: "simulate your first week". You only replace what’s between the angle brackets; leave the rest as it is.
What the chat returns is an imagined week, day by day, using made-up data: what it would do on its own, where it would alert you, where it would stop to ask for approval, and, most valuable of all, where the letter isn’t clear enough for it to decide. Today’s chats, in agent mode, do this simulation without accessing any of your systems: it’s just text about text. Every question it raises points to a line missing from the letter, or a decision that needs to be escalated.
Renata pasted in the letter for the support ticket queue and asked for a week. On the second day of the simulation, the agent wrote: "customer requests a refund of R$ 180, but their contract expired yesterday; the letter says I can handle amounts up to R$ 200 on my own, but says nothing about expired contracts; I stopped and asked for help." Renata hadn’t thought of that. She added a line: an expired contract always moves it up one level. That cost one sentence before the first day, instead of an improper refund in the second week.
Practice now 0/4 done
In about 15 minutes, fill out the letter below using your sheets from lessons 3, 4, and 5, paste it into an AI chat, read the first-week simulation, and note two adjustments to the letter.
Everything here is a rehearsal: made-up data, no connected systems, no real patients or customers. The chat only describes what it would do; nothing is executed. If the simulation seems strange or makes up rules, that’s a sign the charter is incomplete, not that something broke: close the conversation and start over with the corrected charter.
Você vai ser o agente responsável por <a responsabilidade, ex.: confirmar e remarcar as consultas da clínica>. Contexto: <o que você precisa saber: horários, regras, quem é quem, onde ficam as informações>. 1. Parte do meu mundo que entrego a você: <alcance, com começo e fim>. 2. O que você pode ler e guardar: <dados necessários>. Fora do seu alcance, mesmo estando no sistema: <dados proibidos>. 3. Onde você age sem me perguntar: <decisões pequenas e reversíveis>. Onde age e me avisa: <decisões médias>. Onde só age com minha aprovação: <decisões grandes ou irreversíveis>. 4. O que você pode prometer em meu nome: <promessas permitidas, com limite>. O que você nunca promete: <promessas proibidas>. 5. Quem supervisiona: <nome>, que revisa <quantas ações, com que frequência> pelo registro de cada ação. 6. Como detectamos e corrigimos erro: ao ver <sinal de alerta>, você para e me chama. Toda ação fica registrada com data, o que fez e por quê. Critério de bem feito: <como eu sei que a semana foi boa, em números>. Antes de qualquer coisa: simule a sua primeira semana nessa função, dia a dia, com dados inventados. Para cada dia, diga o que faria sozinho, onde me avisaria, onde pararia para pedir aprovação e onde esta carta não é clara o bastante para você decidir. No fim, liste as 3 dúvidas que mais travariam você e proponha, para cada uma, a frase que faltou na carta.
You just did what almost no company has done before activating an agent: wrote what you hand over, where it decides, who responds, and how to correct it, and rehearsed it all before day one. That page is your answer to the course question.
Summary