Module 03 Activity

Scenario

The agent is about to be allowed to do things. This is where you decide what it can never do.

What you produce

A permission specification: actions sorted by reversibility, the narrowest tool for each job, written limits, and a four-category permission list with an enforcement column.

Steps

List every action the workflow could take and write, against each, what undoing it would involve. Mark the irreversible ones.
For each action, describe the narrowest tool that would do the job - what it takes in, what it gives back, and what it refuses. Contrast one with the broad version someone might have proposed.
Write the argument rules for the riskiest action, including at least one limit that is a policy decision rather than a format rule. Say where the limit came from and who owns it.
Describe how a repeated attempt at that action would be recognised as the same attempt rather than a new one, and who would notice if it were not.
Write the permission list in four categories - read, write, act with approval, never - with an enforcement column.
Go down the never list and mark each entry: is it prevented by the design, or only by an instruction? Anything in the second group is not actually on the list.

Evidence to hand in

The reversibility table with the irreversible actions marked.
The narrow tool descriptions, with one broad-versus-narrow contrast.
The argument rules with the policy limit and its owner.
How a repeat attempt is recognised.
The four-category permission list with the enforcement column.
The never list audit.

Review checklist

Actions are sorted by reversibility, not by how significant they feel.
At least one limit is a policy decision with a named owner, not a format rule.
The repeat-attempt scheme is tied to the run and the step, not to a clock.
Every never entry is prevented by the design rather than by an instruction.
No broad 'do anything' tool survives the design.