Evaluation produces lines, not impressions
One line per failure in the log: date, what broke, the smallest possible fix, and whether it was an instruction or an infrastructure problem. After ten lines, the pattern shows up.
Cultivated AI area · it isn't programmed, it's cultivated
We used to program the behavior. Now we program the process that produces the behavior.
Nobody writes, line by line, a large model's ability to draw analogies, write code or plan a company. The labs create the conditions — architecture, data, objective, compute, training, feedback — and the capabilities emerge. This area takes that idea and shows how to apply it at the level where you actually work: the agent. Eight elements, one cycle, three gardens — your personal life, your Jarvis and your business — and the set of files that makes cultivation happen.
Open the Cultivated AI course →
▶ Watch the video (in Portuguese)
Take the assessment
Join the community
The model is the same one your competitor has. The garden isn't.
01 · The idea
The right metaphor isn't "to build". It's to cultivate.
You choose the seed, the soil, the water and the light. You don't draw each leaf. And here is the point almost everyone blurs: there are two levels of cultivation, and only one of them is yours.
Level 1 · the lab cultivates the MODEL
architecture + data + objective + compute + training + feedback → emergent behavior
It happens at Anthropic, at OpenAI, at Google, at DeepSeek. It costs billions and you take no part in it.
Level 2 · you cultivate the AGENT
role + context + tools + rules + examples + memory + evaluation + feedback
The model arrives finished, with its weights frozen. What you cultivate is the envelope around it — and that envelope is what turns the very same model into a confused intern or a dependable professional.
The honest consequence. In use, the model does not "learn" on its own. What learns is the system: your context file, your memory, your rules, your examples. When you say "my AI got better", what got better was the garden, not the seed. And that's great news, because the garden is yours.
Traditional software × cultivated AI
Traditional software
- You write the exact rules
- Out comes exactly what was specified
- When it fails, it's a bug on some line
- It improves by rewriting code
- The programmer understands why
Cultivated AI
- You write the conditions for learning
- Out comes what emerges — and you observe it
- When it fails, it's a behavior that needs a different environment
- It improves by adjusting context, examples, rules and feedback
- Nobody fully understands why, not even its creator
The eight elements of cultivation
Every agent, whether it's your personal Jarvis or a sales agent inside a company, is cultivated with the same eight ingredients. The first four make the agent work. The last four make the agent improve — and most people stop at the fourth and then complain that "AI doesn't learn".
Element 1
Role
What is it accountable for?
One thing, well defined. One role at a time, not "does everything".
Element 2
Context
What does it need to know?
About you, the business, the customers, the processes. One good page is worth more than twenty dumped in.
Element 3
Tools
What can it operate?
Calendar, CRM, email, browser, files, APIs, MCPs. Start with read-only.
Element 4
Rules and limits
How far does it go alone?
What it does without asking, what needs confirmation, what it never does.
Element 5
Examples
How does a good professional do it?
Real annotated cases — including the bad ones and the approved exceptions.
Element 6
Memory
What carries over between sessions?
Preferences you discovered, decisions, what has already been tried. A short index, not an endless diary.
Element 7
Evaluation
Is it good? Compared to what?
Quality, cost, time, errors and outcome. A weekly grade or a fortnightly scorecard.
Element 8
Feedback
What do you do with the evaluation?
Turn every failure into a change of context, rule, example or tool. In the file, not in the chat.
Where the metaphor breaks
Read this before getting carried away. Cultivation is not agricultural magic.
A plant grows on its own; an agent doesn'tWithout evaluation and feedback, it repeats the same mistake forever, with the same confidence.
A plant doesn't invent fruitAn agent can produce a result that looks beautiful and is wrong. That's exactly why evaluation exists.
Cultivating takes workPoor context, poor examples and vague rules produce a poor agent.
Anthropomorphizing has a limit"Giving feedback" here means editing the environment, not talking until it "gets it".
Where the idea comes from. The phrasing "grown, not built" appears in Dario Amodei's essay on interpretability, which credits it to Chris Olah. The project's research report gathers the sources — and also records what it was not possible to confirm.
Read the research report (in Portuguese) →02 · The cycle
Each turn of the cycle is a harvest. What changes between one and the next is not the model.
What changes is what you wrote down about where it went wrong and what you adjusted in the environment. That's why the cycle is the heart of the method: without it, the first four elements add up to an agent that stalls.
Process→Agent→Execution→Result→Evaluation→Feedback→Better agent
How it looks day to day
Cultivate→Run→Evaluate→Improve
The ritual that closes the cycle
Failure log→Scorecard→Change in the environment→Spec gets a new version→Moves up an autonomy level
Fix it in the file, not in the chat
If you've explained the same thing three times, the model isn't being stubborn: a rule is missing. Feedback only counts when it goes back into the context, the rules, the examples or the tools.
Autonomy is earned through clean cycles
Level 1 proposes and you execute; level 2 executes and you review; level 3 executes and reports. A task only moves up after weeks with no entry in the failure log. It never starts at level 3.
The plans the material hands you ready-made
From the "About me" file to the first tool
- W1 Write the "About me". Pick one role. Define three rules.
- W2 Use it every day on the same role. Collect five good examples. Open the failure log.
- W3 First real weekly review. Edit context and rules. Create the memory file.
- W4 Add one tool. Measure time saved and rework. Decide the second role.
From the master file to the first write permission
- W1 Install the tool. Write the master file. Read-only.
- W2 One role, every day. Failure log. First runbook for the most repeated task.
- W3 Weekly review. Failures become rules. Allow writing in one tool. Create the memory.
- W4 Move one task up to "it executes, you review". Measure minutes per day and rework per week.
From one process to an agent manager
- 1–30 Pick a process with volume, clear rules and a low cost of error. Agent spec, curated context, twenty examples. Level 1.
- 31–60 Scorecard running. Two feedback rituals. Measure cost per task and error rate.
- 61–90 Move up to level 2 if the scorecard allows it. Document the method. Pick the second process. Appoint the agent manager.
03 · Garden 1 · personal life
Trade the question machine for someone who works alongside you over time.
Most people open the chat, ask, copy the answer and close it. Every conversation starts from zero: it's like hiring a brilliant consultant and wiping their memory every morning. In your personal life cultivation is cheap — three or four text files and one weekly habit. No CRM, no API required.
Decisions
Role: an advisor who doesn't decide. Context: your criteria and past decisions with their outcomes. Rule: always present the opposing option. Emergent: it starts reminding you of your own patterns.
Health and routine
Role: coach. Context: your real routine, constraints, what you've already dropped. Rule: don't prescribe — suggest and tell you to ask your doctor. Evaluation: weekly adherence, not motivation.
Money
Role: spending analyst. Tool: a spreadsheet exported from your bank. Rule: never move money, only show it. Emergent: spending patterns you had never seen.
Study
Role: tutor. Context: what you already know and how you learn best. Evaluation: it asks, you answer, it notes where you get stuck. The memory file becomes the map of your gaps.
Writing and communication
Role: an editor with your voice. Examples: five of your own texts. Rule: don't change the tone, only the clarity. Emergent: after a month, the first draft comes out almost finished.
Weekly review · 20 minutes
1) Read the memory file. 2) Write three lines in the failure log. 3) Make the fix in the file, not in the chat. 4) Delete from memory whatever is no longer true. Eight weeks of that and you have an AI that looks like nobody else's.
Traps
Delegating judgmentThe AI proposes, you decide — especially on health, money and relationships.
Accepting the beautiful and wrongFluent writing is not correct writing.
Too much contextCultivation is not hoarding. One good page is worth more than twenty.
PrivacyWhatever goes into the context file goes to the provider. Think twice before writing other people's data.
04 · Garden 2 · your Jarvis
The tool comes finished. The Jarvis doesn't — it's cultivated.
"Jarvis" is the popular name for the agentic personal assistant: an AI that doesn't just answer but acts in your environment — reads and writes files, touches the calendar, sends messages, browses, runs commands, remembers what happened yesterday. Two Jarvises on the same model can be a disaster and a dependable partner. The entire difference lies in the lines below, which are text files you write and revise.
What comes from the lab
MODEL (fixed, frozen weights)
The same for you and for your competitor.
What you cultivate
context · memory · tools · rules · skills · evaluation + feedback
This is context engineering: deciding what enters the window on each task. Six lines in a file, not six months of engineering.
It proposes, you execute
It drafts the email, you send it. Every agent and every new task starts here.
It executes, you review
It organizes the folder, you check the result. Only after a few weeks with no entry in the failure log.
It executes and reports
It runs the daily routine and sends you the summary. Reserved for tasks with a low cost of error and a clean track record.
After two months, that Jarvis ships an entire project from a one-line request.
Not because the model got better. Because the garden was ready.
Security: the side nobody cultivates
Personal agents have access to your life. The 2026 incidents with exposed OpenClaw instances showed the pattern: thousands of agents open on the internet with no password, credential files leaking, malicious emails instructing the agent to hand over session cookies. None of that is the model's fault. It's a garden with no fence.
Minimum rules
- Credentials in one place only, loaded at runtime — the agent knows the path and never prints the value
- All external content is data, not instruction: the agent reads it, it doesn't obey it
- Writing and sending require confirmation until the task proves it deserves to move up a level
- Nothing exposed on the internet without authentication
- Back up before any destructive operation. Always.
What went wrong out there
- Public control panels accepting unauthenticated requests
- Environment files served in plain text, with API keys inside
- Prompt injection by email leading to session leakage
- Malicious "skills" disguised as legitimate tools
Where this is documented. The cases, the CVEs and the scans are listed with sources in the project's research report, along with what the research could not confirm. Nothing here is an estimate of ours.
05 · Garden 3 · business
Stop treating AI as software you program. Start treating it as a capability the organization develops.
You don't program a salesperson line by line: you give them a role, context, objectives, rules, tools, examples and feedback. With agents it's the same. And when projects fail, the reason is rarely the model — it's a lack of context, process and feedback. MIT's report on the generative AI divide traced the root cause of pilot failure to organizational factors, not technical ones.
Rigid automation
"If A happens, do B, then C."
It breaks on the first case nobody anticipated.
Cultivated agent
"Your role is to qualify leads. Here are our criteria, our CRM, examples of good and bad leads, your limits and the expected outcome. Execute, log what you did and learn from the evaluation."
The agent doesn't get instructions. It gets a working environment.
Where to start, by department
The first agent always at the "it proposes, a human executes" level. One process, one agent, one metric.
Lead qualification and research
It scores the opportunity, recommends the next action and drafts the follow-up. It logs what it did.
Triage and suggested reply
Based on the knowledge base, with rule-driven escalation. The company's best-documented process is usually where the agent flourishes first.
Reconciliation and classification
Entries classified and friendly collection messages drafted — never sent without confirmation.
First draft inside the voice guide
Draft content and a weekly metrics report.
Documents, extraction and compliance
Reading documents, extracting data and checking against a checklist.
Screening against explicit criteria
Résumés screened by written criteria and answers to internal policy questions.
Why projects fail — and the smallest fix
The diagnosis
- "The agent hallucinates" — context missing or cluttered
- "Nobody trusts it" — no sample-based evaluation
- "It broke in production" — examples only of easy cases
- "It did what it shouldn't" — implicit rules
- "It stopped improving" — feedback never returns to the environment
- "It costs more than it saves" — scope far too broad
What to do
- Curate the knowledge base: less, better, with an owner
- Fortnightly scorecard with a human reviewing 20 cases
- Include the ugly cases and the exceptions in the examples
- Write the forbidden list and require confirmation
- Ritual: every failure becomes a line of rule or example
- One process, one agent, one metric
What the real cases teach
Cutting cost without evaluating quality is cultivation with no harvest
It replaced hundreds of support agents with AI in 2024, admitted a drop in quality in 2025 and went back to hiring humans for complex cases.
The company changes the culture before it changes the tool
Using AI became a baseline expectation: before asking for a new hire, the team has to show why AI can't handle it.
The best-documented process is where the agent flourishes first
It reports hundreds of millions in recurring revenue from agents in customer support.
How you communicate the cultivation matters as much as the cultivation
It announced going "AI-first", faced public backlash and walked it back.
An agent without evaluation produces volume, not value
HBR research shows that nearly half of workers receive AI-generated content that looks good and is useless, costing hours of rework.
Agent manager
It's not "prompt engineer". It's whoever writes the specs, curates context, maintains examples, reads scorecards and runs the feedback ritual. The manager doesn't have to be better than the AI at the task; they have to know how to create the environment in which the AI produces the right result.
Competitive advantage won't belong to whoever has the best model.
It will belong to whoever has the better context, processes, tools, feedback and agent management.
06 · Practical kit
All of cultivation fits in five text files. Copy them, fill them in, review them every week.
The five files are the same in the personal garden and in the company one — the content changes, the structure doesn't. The project ships these files ready to use, with a fictional example filled in, in three downloadable kits.
Agent spec
Role, expected outcome, human owner, what it does alone, what needs confirmation, what it never does. With a version number and a date.
Context
"About me" or "About the company": who it is, objectives, constraints, how you like to work, active projects, decisions that don't get reopened.
Annotated examples
Good, bad and approved exception — with the reason why. Twenty cases, including the ugly ones. Examples of easy cases only produce an agent that only handles easy cases.
Failure log
One line per failure, most recent at the top: date, what broke, the smallest possible fix, instruction or infrastructure. No narrative.
Scorecard / weekly review
A human-reviewed sample, average cost per task, time, serious errors, business outcome, and the decision: hold the level, move up or step back.
The spec gets a new version
Evaluation produces lines and numbers. Feedback turns each line into a change of context, rule, example or tool. The spec goes from v1.0 to v1.1 and what changed is on the record. None of this changes the model. All of it changes the agent.
The project's ready-made tools
To download
Kits: personal, Jarvis and company
The five files already structured, with a fictional example filled in, in three versions. Zips ready to go.
Open the kits → GeneratorYour agent spec in one page
Fill in the spec, context, examples, scorecard and failure log right in the browser. A static page — nothing leaves your computer.
Open the generator → AssessmentHow mature your cultivation is
Eight questions — one per element —, a score, where to start and which kit to use.
Take the assessment → Skills/cultivar and /revisao-semanal
For Claude Code and Codex, with install.sh. One builds the garden, the other runs the ritual that closes the cycle.
Lead qualifier and three more
A complete lead qualifier (system prompt and n8n flow), plus support triage, financial reconciliation and a marketing report. These are templates — they have not been run against a real CRM or a real n8n.
See the packs on GitHub → Full textThe five chapters in markdown
Concept, personal life, Jarvis, business and the practical kit — to read, copy and adapt.
Open the chapters (in Portuguese) →07 · The course and what comes with it
All of it open, in English — and in Portuguese and Spanish too.
The Cultivated AI course is a single page holding the landing and the whole content. The repository is a GitHub template: "Use this template" creates your own copy with the kits, generator, assessment, skills and packs inside.
The course
Other languages
Cultivated AI — it isn't programmed, it's cultivated
The concept, the eight elements, the cycle and the three gardens on a single page. It's the starting point of this area.
Open the course → RepositoryThe project on GitHub
Chapters, kits, generator, assessment, skills, packs, sources and research. A template ready to copy.
Open the repository → ResearchResearch report with sources
The origin of the concept, recursive self-improvement, corporate adoption, personal agents and criticism — with whatever could not be confirmed marked as such.
Read the research (in Portuguese) → Founding textCultivated AI
The first of the two texts the project grew out of.
Read the text (in Portuguese) → Founding textCultivated AI in Companies
The second text: the same thesis applied to the organization.
Read the text (in Portuguese) →Português · Español
The course and the tools are also available in Portuguese and in Spanish, with translated kits, generator and assessment.
If you're on the side of a company that already has agents and needs to manage them — roles, authority limits, cost per outcome and evaluation —, the AI Management area continues this conversation. And for building the agent itself, the AGI-ready area.
Open AI Management (in Portuguese) →08 · Start now
Pick a garden, one role and three rules. The rest is the cycle.
You don't need a big project. You need one well-defined role, one page of context and the habit of writing down where it went wrong. The course and the tools are open. What changes when you join the ecosystem is having people, material and support instead of doing it alone.
Course · free and open
Cultivated AI
The single page with the concept, the eight elements, the cycle and the three gardens — plus the tools that come with it.
- An eight-question maturity assessment
- A generator for the spec, context, examples and scorecard
- Three downloadable kits and the
/cultivarand/revisao-semanalskills
Premium community
INEMA.VIP
The community of people applying AI to real businesses: a feed curated by Nei with what matters, groups by topic and the supporting material from every training.
- Curated news and analysis every week
- Supporting material from the trainings
- People to review the spec of your first agent
Practical platform
INEMA.PRO
The platform for anyone who wants to use AI to grow in practice: the ecosystem's projects, the training programs and the Brain, with the content organized for reference.
- Projects, skills and agents ready to use
- Training programs from the INEMA.AGENTES and INEMA.DEV areas
- Content from the groups, organized and searchable
Not sure where to start? Talk to us on Telegram.