1The problem isn’t creating it, it’s being accountable for it

An agent is an AI that is given a task and carries it out. It can read documents, fill in a spreadsheet, or reply to messages.

The AI Management area starts with an observation. Creating agents has become easy. What holds a company back comes later: who is accountable for the result, what the agent can do on its own, and how to tell if it’s doing well.

Diego discovered that HR already uses three assistants, each set up by a different person. When one of them gave a wrong answer about vacation, no one knew who was responsible.

2Managing an agent is like managing a person

The area’s page says the logic is the same as traditional people management, with more measurement. A person has a job title, responsibilities, tools, limits, and evaluation. An agent has a role, objective, context, tools, autonomy, and evaluation.

There are two new things. The agent needs written context because it doesn’t learn over coffee. And its autonomy is graduated: it gets an authority limit, a limit on what it can decide on its own.

Sônia took the description of the sales associate who handles the shoe store’s WhatsApp. In twenty minutes, she turned it into an agent sheet for customer service.

3Does on its own, confirms first, or never does

The page uses rules the company already knows. In accounting, the person who enters transactions doesn’t approve them. In legal, the agent prepares things, and the lawyer decides and signs.

For each agent, separate the work into three groups: what it does on its own, what it confirms first, and what it never does.

Diego asked an AI chat to sort the tasks for the vacation assistant.

4Don’t speed up the old process

The page’s toughest decision is this: don’t put agents to work carrying out old processes faster. Many steps exist only because there used to be no other way. Speeding up those steps delivers the mistake sooner.

The question for the team is simple. If I had people, agents, and automations today, would this process still exist this way?

Sônia was going to have an agent copy WhatsApp orders into a spreadsheet, as the sales associate did by hand. She stopped first and asked whether the spreadsheet was still necessary.

1From task to responsibility

The page recalls that for three years, we tried to fit the entire operation into one giant request. Today, there are agents that work for hours or days. In that case, you state the destination and the agent finds the path.

The change at your desk is the unit of the request. A task says what to do now. A responsibility says what to take care of, and the agent plans, carries it out, and returns with the result.

Sônia picked up the counter notebook and rewrote three requests she made to AI every week. The list changed its verbs.

2The five parts of a delegation

The page doesn’t say the request is over. It says the request that tries to teach every step is over. The new request has five parts: goal, what matters, what must not happen, available resources, and how you’ll judge the result.

Diego wrote a twelve-step request to prepare the announcement about the health plan change. Every time, a detail was missing. He replaced the steps with the five parts.

3Trust is a curve, not a switch

Delegating doesn’t mean letting go of everything or locking everything down. The page introduces the trust matrix, with three levels: what the agent decides on its own, what it proposes for you to approve, and what it never touches. Each level has an owner.

Sônia created the matrix for the responsibility of restocking inventory.

4What you repeat becomes a manual

The request says what you want now. The way to do it well, which stays the same, goes into a Skill. The page’s tip: if you’ve explained the same procedure three times, turn it into a Skill. In a regular chat, you can start simple: save some text in Notes and paste it along with your request.

Diego realized he was always explaining the same writing rules for HR announcements. He separated the request into two columns.

1The model is the same as the competitor's. The garden isn't

The area opens with a sentence: “Before, we programmed behavior. Now, we program the process that produces behavior.”

The page separates two levels. At the first, the lab cultivates the model, and you don't take part. At the second, you cultivate the agent: the model comes ready, and what changes is the environment around it. The page calls this environment a garden.

Diego and a colleague use the same AI chat to write internal announcements. Hers come out in the company's tone; his are generic.

2The eight elements of cultivation

Every agent, personal or for a company, is cultivated with the same eight ingredients: role, context, tools, rules and limits, examples, memory, evaluation, and feedback.

The first four make the agent work. The last four help it improve. The page notes that most people stop at the fourth and then complain that AI doesn't learn.

Sônia uses AI to write descriptions of shoes for the store's website. She marked what she had in writing.

3Turn every failure into a line

The heart of the method is the cycle: the agent works, you evaluate, the failure becomes an adjustment to the environment, and the agent improves. The page calls each round a harvest.

Two rules support the cycle. Evaluation produces lines, not impressions: one line for each failure, in a log. And the fix goes in the file, not in the conversation, because the conversation gets lost. The file is a text document you keep in Notes or another document, which you paste at the start of each conversation with AI.

Diego opened the failure log for the announcement agent, with the page's four columns.

4Autonomy is earned through a clean cycle

The page defines three levels. At level 1, the agent suggests and you carry it out. At level 2, it acts and you review. At level 3, it acts and reports. Every agent starts at level 1, never at 3.

Everything fits in five text files: the agent sheet, the context, the annotated examples, the failure log, and the weekly review scorecard, where you decide whether to keep it at its level, move it up, or move it back down.

Sônia asked an AI chat to say which level the agent for descriptions should start at.

1The "brain" is what you taught the AI

The area discusses two tools for programming with AI. Claude Code is one. Codex is the other.

The "brain" is everything you’ve accumulated in a project: context, rules, decisions, tasks, and memory. The area advocates keeping this in plain text that any AI can read.

Sônia has been using the same AI chat for months to write the store’s messages. One day, she thought about trying another chat and realized she didn’t know what the first one "knew."

2What keeps you tied down is in the tool, not the model

The page says: "The lock-in isn’t in the model. It’s in what you left inside the runtime." Runtime is the program where the AI runs and stores its memory and history.

The area’s rule is simple: the content lasts; the format is disposable. Mixing the two is what makes switching difficult.

Diego discovered that the rule for writing job postings existed only in the chat’s memory. No one else in HR knew about it.

3Audit, adapt, test, and leave notes for whoever continues

The area’s method has four steps, in this order. Audit means listing what exists, without changing anything. Adapt means separating what works with any AI from what only works in that tool. For example: “exchanges only with a receipt” works with any AI; a button or shortcut that exists only in that app doesn’t carry over.

Test means opening a new conversation and checking whether the AI uses the document. The last step is the handoff: a text with decisions, pending items, and the next step.

Sônia pasted the store document into a new chat and asked a test question.

4Move now or stay independent

The page breaks the decision into three effort levels. Level 1: import into the app itself with one click, limited to what the other side accepts. Level 2: migrate with the area’s kit, a set of ready-made requests that makes the move without deleting anything.

Level 3: the knowledge lives in the project, and the tool becomes just the one that carries out the work. According to the page, it’s the only level that survives the next switch. That doesn’t mean abandoning the AI you use today: you choose the tool for each task.

Diego isn’t switching chats now. Even so, he saved the HR rules in a document so he wouldn’t depend on just one.

MODE: audit
Sem alterar nenhum arquivo, faça um inventário deste projeto.
Devolva só o relatório e um plano. Não implemente nada.

1Writers tend to approve what they wrote

The Codex + Claude area is about two AIs from different companies. Claude is from Anthropic. Codex is from OpenAI, the same company as ChatGPT. The page doesn't ask you to choose one; it asks you to use both, each for a different part.

The reason is simple. If the same AI writes and reviews in the same conversation, it tends to think everything is fine. Two AIs agreeing is not proof either. The test and your approval are what count.

Diego asked an AI for a welcome message for the new hires. Then, in the same conversation, he asked whether the message was good.

2The same request and the actual text

The rule that applies across the board: the second assistant gets the same request the first one got and the actual text it produced. Don’t say, “The other AI wrote something about vacation. Is it good?” Without the original, the review becomes a guess.

Sônia asked an AI for a WhatsApp message telling customers about the winter sale. To review it, she opened another AI and pasted in both things.

3Findings with the smallest fix

A good review doesn’t rewrite everything. It lists findings: what’s wrong, where it is, why it matters, and the smallest fix that solves it. That way, you can decide on each item without losing the text that was already working.

Sônia asked for findings, not a new text.

4Two rounds and a note for later

The page sets a limit: no more than two rounds. In the first, the reviewer points things out; you make corrections. In the second, it checks the corrections. After that, the decision is yours. Endless rounds just waste time.

The last step is a handoff: a short note with the status, decisions, and next action. Whoever picks this up tomorrow, you or another AI, reads it before taking action.

Diego finished reviewing the welcome text with three lines in Notes: "Text approved in round 2. I rejected the finding about the formal tone. Still to do: send it to the manager."

codex exec --sandbox read-only -c model_reasoning_effort=medium \
  --output-last-message review.md \
  'Leia plan-v1.md. Aponte falhas de correção e testes faltando, com evidência, impacto e a menor correção. Não altere arquivos.'

claude -p --model opus --effort medium \
  'Leia plan-v1.md e revise só lendo: falhas, suposições e testes faltando. Não altere arquivos.' > review-claude.md

1From a casual chat to a work system

OSWork presents itself as "AI as a work system" and "from chat to your environment of agents". The idea is to organize files, instructions, and memory so AI doesn’t have to hear everything again in each conversation.

The method has four steps: define the goal, choose the context, do a small task, and check the result.

Every Monday, Sônia asks for the week’s promotion text. And every Monday, she explains the store’s name, tone, and exchange rules all over again.

2All five editions in one table

The page offers "five ways to organize your work with AI" and asks you to choose based on your experience and available time. The same idea is presented in different levels of depth.

According to the page, there are courses and videos in Portuguese, Spanish, and English. v6.2 and v6 have their own links in English and Spanish.

3Two questions determine the edition

The first question: how far do I want to go? The complete path (v2 and v6.2) goes from chat to a consultation bot on Telegram. It ends with supervised operation on a rented computer on the internet. The editions for beginners (v6, v5, and Quick) stop earlier: folders by topic, a sheet with your rules, reusable requests, and a review routine.

The second: How much technical background do I have? v2 uses the terminal and Codex. v6.2 follows the same path in short, visual lessons, for people who have never opened one.

Diego wants a robot in the future that answers the team's questions through Telegram, but he has never used the terminal. He answered both questions.

4Start with a small task

The page says: "Start with a small task." There are three steps: choose where to start; create a practice folder and write a few lines in it about what the task should deliver; save the result, write down a question, and move on to the next lesson.

Two answers in the "Before you start" section ease your worries: you can start with the fundamentals without programming, and you can read the content without signing up. AI subscriptions and renting a computer on the internet are separate services that you choose. And the course does not promise an unsupervised autonomous agent.

Sônia created a folder on her computer called "Treino OSWork". Inside, she saved the week's promotions text and a question.

1Three levels that should not be confused

The RSI area explores a practical question: how an AI system proposes changes, tests the result, and keeps what works. The page separates three levels.

At the first level, AI reviews the response: it critiques and rewrites what it wrote. This can improve a task without changing the model. At the second, it changes the system: the instruction, memory, or tools of the agent. At the third, investigating recursion, the improvement starts helping produce future improvements, and that calls for strong evidence.

A salesperson offered Sônia a virtual assistant that "learns on its own." She wrote down the promise and asked one question for each level.

2Set up the test before making a change

The page suggests starting small. Choose a task, such as pulling action items from meeting minutes. Write the correct answer for three made-up examples.

Then separate the examples into two groups. Two are for adjusting the request. One is reserved: you only use it at the end to check whether the change holds up.

Diego uses AI to extract action items from the HR team’s weekly meeting minutes. Sometimes it forgets who is responsible for each action. Before changing the request, he prepared the test sheet.

3One change, compared using the same examples

Suggest one change to the request. Run the old and new versions on the same meeting minutes. Note the correct answers, errors, and how long you took to review each one.

If you change three things at once and the result improves, you won’t know which one helped. And if it gets worse, you won’t know that either.

Diego’s change was one sentence: "If the meeting minutes don’t say who is responsible, write 'no one assigned' instead of guessing."

4Validate, decide, and make sure you can go back

Only now run the new version on the reserved meeting minutes. If it gets those right too, you can adopt it. Note who decided and where the old version is.

The team calls this a cycle with safeguards. LOOP-R, the cycle proposed by the project, has an explicit rule: no worse version can replace the current one based on a system decision.

Version 2 also got the answers right on Minutes C. Diego adopted the new version, wrote "approved by Diego," and left Version 1 in the same folder with the date.

1An open question turns into a paragraph; a closed question turns into a decision

Jev is TypeSafe’s structured decision model, made by the company TypeSafe. Jev Decision Lab is INEMA’s educational project for learning how to formulate these questions. The central idea: AI receives context and criteria and returns a choice, a probability, or a level. Your process takes action afterward, not AI.

Sônia needs to restock children’s sneakers and is unsure which of two orders to place with the supplier. First, she asked the usual way.

2Three ways to ask

The page presents three ways. The right one depends on the type of answer you need. Noul is the name for a yes-or-no question: instead of just “yes” or “no,” it returns the chance that the answer is yes.

Diego receives dozens of employee requests each week: schedule changes, letters, reimbursements. “Which queue should this request go to?” is Choice. “Does this request have all the documents?” is Noul. “What is the urgency, from low to high?” is Score.

Choosing the wrong format gives you an answer that’s hard to use: a score when you needed a queue, or a yes when there were several levels.

3The policy: proceed, abstain, or send it to a person

A choice alone is not enough. You also decide in advance what to do with it: when the suggestion proceeds, when AI should say “I don’t know,” and when the case goes for human review. The page cautions: confidence is not automatic authorization.

Diego set up the Choice question for the requests and added the policy.

4What has been observed and what still needs to be measured

This is the section that makes the area trustworthy. Each number comes with what it proves and what it doesn't prove.

Sônia liked the AI's choice between requests A and B. Before using it every week, she noted: "Already observed: the choice matched mine in one case. Still needs to be measured: about ten restock orders, compared with what I would decide."

1The agent guesses, or the site tells it

An agent visiting a site today does what a person would do: looks for the field, types, clicks the button, and hopes it worked. If the layout changes, it gets lost.

WebMCP is another way. The page itself tells the agent what it can do: each action has a name, a description, and what it requests. The agent calls the right action by name. The area’s page says the technology is still experimental and available for testing in the browser.

Sônia noticed that a customer asked an assistant whether the store had a certain model. The assistant would have to open the site and search on its own.

2The free assessment shows where you are

Before studying, the area asks you to measure. WebMCP Readiness, at webmcp.inema.pro, opens your site’s address in a disposable browser and only observes. It does not trigger anything on your site.

You get an overall score and four separate scores: WebMCP, SEO, GEO, and AEO. The last two measure whether the site can appear in and be cited in AI assistant responses. Then come the blockers and a plan of up to 12 fixes in order.

3Six fixes for this week

The page lists six fixes the site owner makes before any phase. Each one is tied to an item the assessment checks. The method is short: fix it, run the assessment again, and watch the score go up.

Diego opened the list thinking about the company website, where the jobs page is. He wanted to know which item to ask the team that manages the site to address first.

4The diagnosis doesn’t prove everything

The page is clear about its limits. The diagnosis shows what can be observed from the outside. It doesn’t promise that the site will rank in search results or that an AI will cite it.

And there’s one rule that applies across the whole area: AI chooses the action, but the site decides whether it can happen. A person still needs to check who can view a request and what requires confirmation.