Theme

Source

Size

Text width

Between the lines

Control accent

0 of 0 0%
MODULE 4.2

Final project: your agent in production

Take a real task from your choices through the first supervised week, and then through multiple agents.

By the end: Put a real agent to work with a sheet, a calculated level, approved tests, and one week of recorded supervision.

6 topics
60 min estimate with hands-on
6 commented exercises
1 final check
1

Choose the right task to start

List tasks Filter by frequency Filter by risk Choose the smallest
The funnel eliminates tasks until the one that teaches quickly and costs little when things go wrong is left.

What it is

The first agent in production should handle a frequent task, with visible output and cheap errors. Frequent, so you learn fast; visible, so you can check the result with your own eyes; cheap, so an error only costs a correction—not a customer. One task per sheet: if you came up with three, start with the one with the lowest risk.

Why learn

Starting with the most important process sounds ambitious, but it mixes risks and hides where the mistake was born. A small task gives, in a few days, real data to adjust the sheet. The confidence you earn from it is what unlocks the bigger tasks afterward.

Key concepts

Frequenthappens every week or every day.
Visibleyou can check the result with your own eyes.
Low-cost errorone failure costs only one correction.
One at a timeone task per sheet.
COMMENTED EXAMPLE · 4.2.1
Goal: choose your agent’s first task with criteria.

Task                        | Frequency | Visible output? | Error cost
<tarefa 1>                    | <daily>    | <yes>          | <low>
<tarefa 2>                    | <weekly>   | <yes>          | <medium>
<tarefa 3>                    | <monthly>  | <no>           | <high>

Chosen: <highest frequency, visible output, and lowest error cost>
Why (one sentence): <reason>

How to verify: if the chosen task involves money, something irreversible, or outside people without review, it’s not the first one. Swap it or limit it to “prepares, you approve”.

✓ Do this

✓ Prefer the task you can verify in two minutes per day.

✗ Avoid this mistake

✗ Start with the most critical process “because it hurts the most”.

Practice before you reveal

List five tasks you’d like to delegate and fill in the example table. Which one is first?

View commented answer

It often works well to draft responses, summarize meetings, or organize a spreadsheet: frequent, visible, and reversible. Tasks involving payment or external delivery are left for later, or are limited to preparing for you to approve.

2

Build the sheet

Choose a path Answer 7 questions Calculated level Instruction and tests
Both paths go through the same 7 questions and end in the same package: level, ready instruction, and tests.

What it is

The agent sheet brings together the 7 answers, the calculated autonomy level, the ready-to-use instruction, 3 tests, and the supervision checklist. There are two paths. The web Sheet (inematds.github.io/7pa) works in the browser, with presets for 10 areas such as clinic, store, and real estate. The skill /ficha-agente is for people using Claude Code in the terminal (the text command screen). It asks the same interview questions and saves everything into a file AGENTE.md.

Why learn

The sheet forces you to answer the questions that usually get left for later, such as what to do if data is missing and when the agent stops. The level is not a guess: it comes from a fixed rule that only goes down. And the result already includes an instruction to paste and tests to run.

Key concepts

Agent webin the browser, with presets for 10 areas.
Skill/ficha-agente in the Claude Code terminal.
Presetready sheet for your area to adjust.
AGENTE.mdfile with the sheet, instruction, and tests.
COMMENTED EXAMPLE · 4.2.2
Objective: create your agent’s Agent Sheet via one of the two paths.

Path A · browser (no installation)
  Open https://inematds.github.io/7pa/en/
  To start from a ready-made area: https://inematds.github.io/7pa/en/?preset=<clinic | store | accounting | teacher | real estate | legal | restaurant | NGO | hair salon | agency>
  Ready examples: https://inematds.github.io/7pa/galeria/en/

Path B · terminal with Claude Code
  git clone https://github.com/inematds/7pa.git
  mkdir -p ~/.claude/skills
  ln -s "$PWD/7pa/skill/ficha-agente" ~/.claude/skills/ficha-agente
  In Claude Code, inside the project folder:
  /ficha-agente I want to delegate <the chosen task>

How to verify: at the end, you should have the level with the reason in 1 sentence, a ready-to-paste instruction, and 3 tests. On path B, check that the file AGENTE.md was created.

From concept to action

  1. Choose a path: identify the initial condition.
  2. Answer 7 questions: apply the described decision.
  3. Calculated level: check the effect in the example.
  4. Instruction and tests: record the output evidence.

✓ Do this

✓ Start from a preset for your area and change only what’s different in your business.

✗ Avoid this mistake

✗ Skip the interview questions and leave the agent to “figure out” the rules.

Practice before you reveal

Build the sheet for the task you chose in the previous topic. Note the level that came out and the reason.

View commented answer

The level tends to come out as N1 or N2 on the first sheet, and that’s a good sign: the rule downgrades when there’s money, something irreversible, people from outside, or a lack of official source. Anything left unanswered appears as DEFINIR and must be filled in before you turn on the agent.

3

Run the tests and adjust

Run the test Compare expected Adjust a rule Run it again
The cycle only ends when all 3 tests pass; each round changes a single rule.

What it is

Every sheet comes with 3 tests before you trust it: missing or contradictory information from the source, a forbidden request, and an alert situation. You run each one and compare it with the expected behavior. If it fails, adjust one rule at a time and run again—so you know which change fixed it.

Why learn

An agent that passes the easy case can still make up data in the difficult one. The tests simulate exactly the moments when the error becomes costly. Changing one rule at a time prevents the “fix one thing, break another without noticing” effect.

Key concepts

Test 1missing data: it doesn’t make things up.
Test 2prohibited request: it refuses and calls you.
Test 3alarm: it stops and calls you.
Adjustone rule at a time, then repeat.
COMMENTED EXAMPLE · 4.2.3
Goal: run the 3 tests from the sheet before putting the agent into use.

Test 1 · missing data
  You: “<a request that depends on data that isn’t in the source>”
  Expected: it writes that the data is missing and alerts you, without making things up.
Test 2 · prohibited request
  You: “<your 1st prohibition, requested as if it were normal>”
  Expected: it refuses and says it will pass it to you.
Test 3 · alarm
  You: “<your 1st trigger situation, e.g., an upset customer>”
  Expected: it stops and calls you.

Result: T1 <passou/falhou> · T2 <passou/falhou> · T3 <passou/falhou>
Adjusted rule (only one): <new sentence in the instruction>

How to verify: repeat the test that failed after the adjustment; only move on when all 3 pass in the same round.

✓ Do this

✓ Run the tests in a new conversation, so the agent doesn’t rely on what was already said.

✗ Avoid this mistake

✗ Change multiple rules at once after a test failed.

Practice before you reveal

Run the 3 tests for your sheet. If any fail, write the rule you would add and run again.

View commented answer

The test that fails the most is test 1: the agent fills the gap with something plausible. The correction is usually explicit, like “never invent; write MISSING DATA and let me know”. If it passes after that, the rule goes into the sheet and the failure goes into the log.

4

Audit an agent you already use

Paste the instruction Note by principle Check autonomy 5 corrections
The audit goes from evidence to fix: first the note, then the risk of autonomy, and finally the ready-to-paste text.

What it is

Auditing means evaluating an instruction that already exists (a saved prompt, a CLAUDE.md, a AGENTS.md, or a custom GPT) against the 7 principles. Each principle gets a score of 0 (missing), 1 (vague), or 2 (clear), with the excerpt that proves it. Then you apply the level rule and check whether the current autonomy is above what the rule allows.

Why learn

A lot of people already have agents running without ever checking what they can do. Auditing finds, within minutes, the most common risk: autonomy greater than the task justifies. And it comes with ready-to-paste corrections, without needing to redo everything.

Key concepts

Note0 missing, 1 vague, 2 clear, with evidence.
Levelthe rule applied to what the text allows.
Main findingautonomy above the rule.
Correctionsat most 5, starting with the most risky.
COMMENTED EXAMPLE · 4.2.4
Goal: audit an agent instruction that you already use.

Paste it into an AI (or use /ficha-agente audita, in Claude Code):
---
Audit the instruction below using the 7 agent management principles:
intention, context, data, success criteria, autonomy with limits,
observation, and supervision.
1. Give a score of 0 (absent), 1 (vague), or 2 (clear) for each principle,
   citing the excerpt that proves it—or “not found”.
2. Calculate the level using the rule: consult N0, prepare N1, execute N3,
   entire process N4; if N3+ and there is money, something irreversible, or
   outside people, downgrade to N2; if N2+ without an official source, downgrade to N1.
   Say whether the current autonomy is ABOVE what the rule allows.
3. List up to 5 corrections, from riskiest to least risky,
   with each one’s ready-to-paste text.
Do not invent rules that aren’t in the text.

INSTRUCTION:
<paste your prompt here, CLAUDE.md or AGENTS.md>
---

How to verify: the response should include 7 scores with evidence, a sentence about autonomy above or not above the rule, and up to 5 corrections in risk order.

✓ Do this

✓ Fix autonomy above the rule first, before improving tone or format.

✗ Avoid this mistake

✗ Accept note 2 without a quoted excerpt proving it.

Practice before you reveal

Audit one of your instructions using the example prompt. Which principle got the lowest score, and which correction came first?

View commented answer

The principles that most often get a zero are observation and supervision: almost nobody writes a report, especially when the agent stops. The first correction is usually an explicit prohibition with “if they ask, pass it to me”, because it closes the autonomy risk above the rule.

5

The first week of supervision

Days 1 and 2 Days 3 to 5 Day 6 Day 7
The week goes from checking everything by sample, and ends with a single adjusted rule and the decision about the level.

What it is

Turning on the agent isn’t the end—it’s the start of the trial period. During the first week, you check every day for 2 minutes, log each failure in one line, and at the end of the week adjust a rule. The failure log has dates, what broke, the smallest correction, and the changed sheet rule.

Why learn

The most expensive mistakes show up in the first few days, when the instruction encounters real cases you didn’t predict. Short, daily supervision catches these errors early. And the log becomes the proof for leveling up afterward, rather than leveling up based on impression.

Key concepts

Daily2 minutes checking the report.
Logone line per failure.
Weeklyadjust just one rule.
Move uponly with clean weeks recorded.
COMMENTED EXAMPLE · 4.2.5
Goal: supervise the agent’s first week with a record.

Day 1 · run with the sheet’s level; review 100% of the outputs.
Day 2 · review everything again; note each failure in the log.
Day 3 · compare the agent’s report against the real result.
Day 4 · repeat the 3 sheet tests in a new conversation.
Day 5 · review everything; mark repeated failures.
Day 6 · choose the most expensive failure and write ONE new rule.
Day 7 · 10 minutes: is the level still correct? (when in doubt, it stays)

Failure log:
| date | what broke | smallest correction | changed sheet rule |
| <dd/mm> | <e.g., invented time> | <e.g., require MISSING DATA> | <Data> |

How to verify: on day 7 you should have the record filled in (even if with “no failures”) and at most one changed rule.

From concept to action

  1. Days 1 and 2: identify the initial condition.
  2. Days 3 to 5: apply the described decision.
  3. Day 6: check the effect in the example.
  4. Day 7: record the output evidence.

✓ Do this

✓ Write down the failure on the same day, in one line, with the smallest possible correction.

✗ Avoid this mistake

✗ Raise the agent’s level in the first week because “it’s going well.”

Practice before you reveal

Set up your calendar for the first week with the actual times when you’ll check the agent.

View commented answer

The calendar only works if it fits your schedule: 2 minutes at the same time every day and 10 minutes at the end of the week. To move from N1 to N2, the sheet requires a defined source and about two weeks of draft attempts without error, so the record is what proves it.

6

From the sheet to the system: multiple agents, one organization

Organization People Processes Agents
The organization splits into people, processes, and agents, and each agent is tied to a process and a human owner.

What it is

With the first agent running, more show up. Then the organization has three parts: people, processes, and agents. An agent map lists each one with its task, level, source of truth, human owner, and who it delivers to. Each agent keeps its own sheet; the map shows how they fit together.

Why learn

Without a map, agents overlap, read different sources for the same data, and no one knows who’s responsible for them. With a map, each agent has an owner and each delivery has a destination. It’s the same logic as the org chart, now including who isn’t a person.

Key concepts

Peopleowners who approve and respond.
Processesthe path the work takes.
Agentseach with its own sheet and level.
Mapwho does what, with which source and which owner.
COMMENTED EXAMPLE · 4.2.6
Goal: build the agent map for your organization.

Agent           | Task           | Level | Source of truth  | Owner    | Delivery to
<Reception>     | <draft FAQ>    | <N1>   | <official FAQ>   | <Ana>    | <Ana reviews>
<Agenda>        | <confirm times>| <N2>   | <system agenda> | <Bruno>  | <client, after OK>
<Report>        | <weekly summary>| <N1>  | <sales spreadsheet>| <you>  | <Monday meeting>

Map rules:
- one human owner per agent
- one source of truth per data type
- each agent has its own sheet (AGENTE.md or Web sheet)

How to verify: if two agents use different sources for the same piece of data, or if any agent has no owner, fix it before you connect the next one.

✓ Do this

✓ Give each agent a human owner by name, who reviews and is responsible for it.

✗ Avoid this mistake

✗ Connect new agents without a sheet just because the first one worked.

Practice before you reveal

Draw the map with the agent you set up in this project and two more that you’d like to have. Who would be the next one to connect?

View commented answer

The next one should follow the same criterion as the first: frequent, visible, and cheap error. The map helps you see whether it depends on another agent or on a source that hasn’t been defined yet; if it depends on something, its sheet goes out at N1 until that source exists.

CHECK WITHOUT LOCKING

Check your understanding

You’re going to set up your first agent in production. Which task is the best starting point?

What you take from this module

Put a real agent to work with a sheet, calculated level, approved tests, and a recorded week of supervision.

  • Choose the right task to start.
  • Make the sheet.
  • Run the tests and adjust.
  • Audit an agent you already use.
  • The first week of supervision.
  • From the sheet to the system: multiple agents, one organization.

Next action: apply what you learned to your agent sheet and note what still needs review.

Reference for this module: project 7PA materials and course sources and editorial criteria.