AGI‑ready area · courses, projects and tips
For three years we tried to fit the entire operation into one giant prompt. With agents that work for hours or days, the game changes: you state the destination, the agent finds the path. This area gathers what’s changing, the mental model to keep up, and INEMA’s courses and projects so you can start today without coding.
01 · What’s changing
Six changes already underway. The first three come from reports of those inside the labs; the next three are what it means on your desk.
An agent receives a responsibility, plans, executes, monitors and returns with the result. According to the video below, within OpenAI the “agentic work days” already exceed those of human researchers, and the stated goal is a fully automated AI researcher by around March 2028.
OpenAI’s chief scientist, Jakub Pachocki, writes that he expects to see “almost fully autonomous” companies, with most of the workload performed by Astra‑class models. Those who delegate well multiply their own capacity; those who only execute compete with the agent.
The same text admits that alignment research and monitoring are lagging behind capacity, and calls for coordination among labs. For you, the lesson is practical: powerful agent without authority, approval and verification is risk, not productivity.
You stop asking "make this spreadsheet" and start delivering "handle this month’s reconciliation". The work unit you delegate levels up, and what you need to know how to write changes along with it.
What the agent decides alone, what it proposes and you approve, what it never touches. Each line has an owner. Without this matrix, either you lock everything down or let too much go.
Define the goal, choose what matters, state what cannot happen and judge if the result is good. That’s the role of an agent manager, and it’s trained.
Source of this block: the Wes Roth video on the article "Alien Minds" by Jakub Pachocki, OpenAI’s chief scientist, and the company’s internal research acceleration report. In English, 29 minutes. It talks about recursive self-improvement, goal alignment versus value alignment and the request for coordination among labs.
▶ Watch the video (English)The other source: the video "AGI has arrived", about the launch of GPT-6 Astra and the arrival of super-agents, with the cases of Ethan Mollick, the Ship Closer agent from Vercel and the 41 financial documents vetted in a round. He turned the course The Super-Agents Have Arrived, in 6 lessons.
▶ Watch the original video02 · It’s not the end of the prompt
Talking about the "end of the prompt" is an exaggerated simplification. The prompt does not disappear. What is ending is something else: between 2023 and 2025 we practically programmed the AI in natural language, putting all the operation’s intelligence into a gigantic prompt. With more capable agents, the prompt levels up: moves from "how you should do it" to "what I want to happen".
Before · micro-prompting
"Research five competitors. Visit their sites. Analyze prices. Create a table. Compare the differentiators. Then write a report..."
A sequence of commands. You teach each step: do A, then B, then C, then D. If a step is missing, the result comes out wrong.
Now · guidance, intention and delegation
"Analyze our competitive position in this market and tell me where the three best opportunities are. Use the available sources and tools. Do not make recommendations without evidence."
It’s still a prompt. But now it states the goal, what matters, the constraints, the resources and what counts as a good result. The agent discovers a larger part of the how.
The better the agent, the less we need to teach the path and the more we need to explain the destination.
Imagine a company. You go to an employee and say "prepare the proposal for client ACME". That’s the prompt. But the company has a manual with the proposal template, the commercial policy, the maximum discount, how to calculate margin, the visual standard, the approval steps and the final checklist. That’s a Skill.
The guidance for that specific situation. It changes with each request. It is the briefing.
Knowledge or procedure reusable: instructions, reference files, templates, scripts and tool usage guidelines. Closer to an operational manual than to a prompt.
You say: "Turn this research into a presentation for Brazilian entrepreneurs." That’s the prompt. The presentation Skill provides the narrative structure, slide template, text rules, visual criteria and the review process. The agent receives PROMPT + SKILL and performs much better. The user doesn’t need to write the nine steps each time.
/skills, and you create one by asking for something like "create a Skill to turn research into lessons in the INEMA standard". In Claude Code, the equivalent is the project's skills folder, which several projects below already provide ready-made.03 · The architecture of a working agent
Each level answers a different question, and the example is the same from start to finish: increase a course's conversion.
Why?
"I want to increase the course conversion."
What?
"Find out why people reach the page and don’t purchase and propose improvements."
Guidance for this execution
"Analyze page, offer, objections, competitors and audience behavior. Prioritize changes that can be tested quickly."
How to do it well, repeatedly
auditoria-de-conversao: analyze offer, identify ICP, study competitors, evaluate headline, check social proof, identify objections, propose experiments, rank by impact, validate result.
What it needs to know
Company, products, customers, history, brand, competitors, policies, objectives, previous decisions.
The agent’s hands
Browser, Gmail, calendar, CRM, database, GitHub, spreadsheets, files, APIs, MCP.
Accumulated experience
What has been tried, previous decisions, preferences, customers, results, errors, exceptions.
Quality control
Permissions, limits, human approval, logs, guardrails, evals, success criteria. We don’t just want a powerful agent: we want an agent controllable.
You say "launch my new course". Without context and without skills, it's impossible to do well. But if the agent has these five Skills, a short phrase like "launch this course for Brazilian entrepreneurs and aim for 200 enrollments" triggers a huge amount of work. That's the leap.
It's similar to hiring a person. When you hire an experienced marketing director, you don't say "click Chrome, open Google, type competitors". You say "we need to increase enrollments for this product by 30%", because the professional already has skills, knows processes, knows how to use tools and has experience. With agents, we are heading in the same direction.
How to ask better.
What information to deliver.
What reusable capabilities the agent should have.
How to make the agent work.
How to coordinate multiple agents, tools, and humans.
BEFORE YOU PROMPTED THE TASK.
NOW YOU GUIDE THE AGENT.
04 · Tips to become AGI-ready this week
Take your list of requests to the AI from last week. Mark what is a task with a recipe and what is a problem that an agent would solve on its own if it had a goal, constraints, and a done criterion.
Replace the sequence of steps with five elements: goal, what matters, what must not happen, available resources, and how you will judge the result. Ask for a plan before any action.
If you explained the same procedure three times, it deserves to become a Skill: manual, template, checklist, and examples. Next time, the prompt stays short.
Three columns: decides alone, proposes and I approve, never touches. An owner with a name on each line. That's the difference between delegating and abandoning.
Who the company is, who the client is, what has already been decided. An agent with many tools and little context moves quickly in the wrong direction.
Success criteria, examples of acceptable results, and a way to verify. Without eval, you only discover the error after it has already cost you.
05 · AGI-ready Courses
All open, in Portuguese, directly in the browser. The first three are for decision‑makers who don’t code. The others use the tool (Claude Code, MCP, skills) for those who want to build.
What changes in your work when AI stops waiting for orders. Task vs responsibility, delegation in an AI chat, trust levels, human competencies, and the delegation letter with six questions.
Open the course → In practice · 8 lessonsTake a real responsibility from your area and hand it to an agent with identity, memory, tools (MCP), authority, autonomy matrix, and trust sheet. The same agent grows lesson by lesson.
Open the course → Processes · 8 lessonsRedesign processes for humans to lead and agents to execute. For those who manage a area and need to decide what leaves people’s desks.
Open the course → Immersion · 4 daysFrom task to intention. Day 4, "The Turn", covers models that operate on intention (Fable 5.1 and GPT‑6 Astra).
Open the immersion → Tools · 2 tracksFive flows for AI to operate your computer. It’s the "tools" level of the architecture, seen up close.
Open the course → Orchestration · 4 tracksFundamentals, hands‑on creation, model, cost, and orchestration. For those who already delegate and want to coordinate multiple agents.
Open the course → Methodology · 10 modulesFrom brainstorming to deployment with agents: TDD, sub‑agents, systematic debugging, and skill creation. Skill Engineering at its core.
Open the course → Coaching · skillA coach who builds with you, layer by layer, Identity, Substrate, Rules, Skills, Tools, and Agents.
Open the guide → Step by step · 6 layersStep‑by‑step agent creation system, no coding required. The map of the six layers in skill form.
Open the guide →06 · Hands‑on Projects
Each project below is a piece of the architecture working for real. Open the guide, copy, and adapt.
Seven models that become your agent’s specification on a page, with a simulated trial before you fire anything up.
Open the guide → Agent with human gateTelegram gateway with durable queue: stateful skills and flows, human approval, and resume. The trust matrix in code.
Open the guide → Chained skillsA link turns into a video: research, script, voice, editable scenes, and MP4, with human approval before heavy production.
Open the guide → Tools for agentsTest if your site is ready to talk to agents. The "tools" side of the architecture, seen from the site. The full training is here in the events, in WebMCP.
Test the site → Context and rulesAnalysis and practical sheets on how a frontier model is guided from within. Reference for writing your own rules.
Open the guide → Research skillVerified multi‑perspective research as a Claude Code skill, with a contradictions map and citation checking. Includes the skill for download.
Open the course →07 · Start now
The courses are open. What changes when you join is having people, material and support instead of learning alone.
The portal with all INEMA courses and projects: tracks organized by level, search by topic and the latest updates. It's where you start.
The community of those applying AI to real businesses: feed curated by Nei with what matters, groups by topic and support material from all trainings.
The platform for those who want to use AI to grow in practice: the more than 400 ecosystem projects, the trainings and the Brain with content organized for reference.
Unsure where to start? Talk to us on Telegram.
