LESSON 03 · AUTHORITY
The Agency Budget
Autonomy is not one switch. Allocate it by action class, reversibility, consequence, and evidence quality.
Give the agent exactly the authority it needs.
“Human in the loop” is too vague for production design. A stronger model assigns authority separately across action classes. An agent may have broad read and analysis authority while having almost no commitment, spending, or deletion authority.
D→V RULE
Agency should increase slower than consequence. As actions become harder to reverse or more externally consequential, the approval requirement should become stronger.
| Action | Meaning | Authority level | Current risk |
|---|---|---|---|
| Read | approved operational data | LOW | |
| Analyze | form hypotheses and reconcile evidence | LOW | |
| Recommend | propose bounded actions | LOW | |
| Modify | change internal reversible state | LOW | |
| Commit | create external operational commitment | LOW | |
| Spend | create financial obligation | LOW | |
| Communicate | send external messages | LOW | |
| Delete | remove records or resources | LOW |
INTERACTIVE LAB · BUDGET THE AUTHORITY
Spend authority where it creates value.
For the freight-exception agent, allocate a reasonable operating budget. The meter rises as you grant broader autonomous authority.
SCENARIO
Freight exception investigator
The agent can investigate late shipments and recommend recovery options. Booking changes, customer promises, credits, and deletion have real external consequences.
Recommended patternHigh read / analyze
ModerateRecommend
LowCommit / spend / delete
AGENCY BUDGET USED
0 / 32
No authority allocated yet.
Read + Analyze0
External consequence0
Recommended next checkAllocate deliberately
LESSON ARTIFACT
Save the Authority Matrix.
The matrix becomes a concrete policy input for tool permissions, approval gates, runtime enforcement, and test cases.