All articles

AI agents

An AI agent needs a permission model before a prompt

The prompt describes desired behavior. The permission model defines what the system can actually see, change, spend, send, publish, or escalate.

A practical approach to defining tools, data boundaries, approvals, and recovery before an agent enters a real workflow.

01

Inventory consequential actions

List every tool action and classify its consequence: read, draft, recommend, modify, send, publish, spend, delete, or grant access. A broad tool that can perform several classes should be split or wrapped so authorization can be enforced at the operation level.

02

Bind permissions to context

Access should reflect the user, account, campaign, document, environment, and current workflow state. The agent should receive the least context required for the task and should not be able to substitute an unapproved target merely because a tool accepts arbitrary identifiers.

03

Place approval before the irreversible step

Human review is useful when the reviewer can see the proposed action, source evidence, affected target, and expected consequence before execution. An approval after an email was sent or a budget was changed is an audit event, not a control.

04

Trace decisions and design recovery

Record the tool selected, validated arguments, authorization result, output, and resulting state without storing unnecessary sensitive data. Define idempotency, timeouts, retries, compensation, and escalation so a partial failure does not silently duplicate or corrupt work.

References

  1. Agents SDKOpenAI Developers
  2. Using toolsOpenAI Developers

This is an internal engineering framework informed by the OpenAI agent and tools documentation. It is not a security certification or complete threat model.