Before deployment

Define:

Agent purpose

What specific job does the agent perform?

Inputs

What information does it receive?

Tools

Which systems can it access?

Actions

What can it actually execute?

Boundaries

What must it never do?

Agent configuration

Purpose
↓
Instructions
↓
Knowledge
↓
Tools
↓
Permissions
↓
Escalation

Purpose
↓
Instructions
↓
Knowledge
↓
Tools
↓
Permissions
↓
Escalation

Purpose
↓
Instructions
↓
Knowledge
↓
Tools
↓
Permissions
↓
Escalation

Purpose
↓
Instructions
↓
Knowledge
↓
Tools
↓
Permissions
↓
Escalation

Permission model

Read

Can retrieve information.

Write

Can modify information.

Execute

Can perform actions.

Approve

Requires explicit authority.

Do not automatically grant all four levels.

Evaluation

Test the agent against:

• Expected tasks

• Ambiguous requests

• Incorrect information

• Missing information

• Unauthorized requests

• Tool failures

• Escalation scenarios

Production checklist

• Purpose defined

• Owner assigned

• Tools verified

• Permissions restricted

• Evaluation completed

• Monitoring enabled

• Escalation configured

• Rollback path documented

Production readiness is determined by behavior under failure—not only performance under ideal conditions.