Cultivated AI|INEMA.CLUB-PRO

Lesson 7 · 16 min

Autonomy by evidence, not by trust

The temptation is to grant autonomy early: the answers look good and someone says "just let it run on its own." It's like promoting the intern on day two because they asked a good question. This lesson gives you the four levels, the numbers to move up, and the way back.

Start the lesson
Driving school practice track seen from above, with a car and orange cones.

In this lesson you will

  1. 1The four levels: recommends, executes with approval, executes with exceptions, operates within limits
  2. 2The numbers to move up a level
  3. 3When to step down, and why it isn't punishment
  4. 4Practice the conversation with someone who wants to skip steps

Goal: decide when to raise or lower an agent's autonomy level using accuracy, run volume, and the severity of errors.

01The four levels of autonomy

One task at a time. Click each level to see the criteria for moving up.

The same ladder applies to everything: personal life, Jarvis, and your business. The kits, skills, and project packages use these four levels.

The criterion for moving up isn't a suggestion: it's a requirement. Without the numbers, the promotion doesn't happen.

02What to measure before promoting

Eight metrics. The first three are the minimum.

1. Volume and accuracy

History of runs in that scenario and accuracy rate by evaluation (human or automated).

03Simulated conversation: the director in a hurry

A simulated conversation for practice. Pick the answer you would give; the comment appears right after.

Sales directorwants the leads agent sending emails on its own

Choose your answer. The comment appears right after, and the conversation continues based on your choice.

    04Promotion criteria

    Copy this into the agent profile, one table per task.

    In FICHA-DO-AGENTE.md
    ## Autonomy — task: <e.g.: send follow-up>
    Current level: 1 (recommends; human decides and executes)   since: <date>
    
    To move up to level 2 (executes with approval):
    - [ ] 50+ evaluated runs        today: 
    - [ ] accuracy above 90%            today: 
    - [ ] zero serious errors in 2 weeks   last serious error: 
    
    A serious error on this task is: <e.g.: wrong price, deadline, or recipient>
    Moves down one level if: a new serious error appears or accuracy drops below <x>%
    Next review: <date>

    05Practice

    One task, one criterion.

    1 · The criterion for your task

    15 min

    Pick one task for your agent (personal, Jarvis, or company) and fill in the promotion criterion. Define what a serious error is for it.

    Done when: the profile has the current level, the three numbers to move up, what a serious error is, and the rule for moving down.

    2 · Simulated conversation with a top score

    10 min

    Do this lesson's simulated conversation. If you don't reach the top score, read the comments and start over.

    Done when: you finished with the top score and you can say why "wrong price" kept the agent at level 1.

    06Quick quiz

    4 questions.

    Question 1 of 4

    An agent has 120 evaluated runs at level 1, 93% accuracy, and no serious errors for three weeks. Can it move up?

    07Summary

    Four levelsRecommends, executes with approval, executes with exceptions, operates within limits.
    Numbers to move up90%/50, 95%/200, 98%/500, always with zero serious errors in the period.
    Moving down is part of itNew error, one level down, until it stabilizes.

    Going further

    Previous lessonNext lesson