AI Agent Safety 101: Permissions, Approvals, and What Not to Automate

Some actions should always require human approval:

• Money movement — payments, transfers, purchases.
• Identity changes — password resets, account modifications.
• Legal commitments — contracts, agreements, compliance filings.
• Medical decisions — diagnoses, prescriptions, treatment plans.
• Irreversible actions — data deletion, public statements, account closures.

If in doubt, add it to the red lines list. You can always remove restrictions later.

Structure agent access in tiers:

1. Read-only: The agent can view data but not change anything.
2. Draft mode: The agent prepares outputs but doesn't send or execute.
3. Supervised execution: The agent acts but shows a preview for approval.
4. Autonomous execution: The agent acts independently within defined scope.

Start at tier 1 or 2. Move to higher tiers only after consistent performance.

An approval gate is a checkpoint where the agent must:

1. Show a preview of the action it wants to take.
2. Wait for explicit confirmation before proceeding.
3. Log the approval decision for audit.

Use approval gates for: sending messages, modifying records, scheduling meetings, and any action that affects other people.

Agents should escalate to a human when:

• They encounter a situation not covered by their rules.
• Confidence is low on a critical decision.
• The user explicitly asks for human involvement.
• A red-line action is triggered.

Good escalation includes context: what the agent tried, what it's uncertain about, and what it recommends.

Every agent action should be logged:

• What was accessed or modified.
• When the action occurred.
• Why — the trigger or decision logic.
• Who approved it (if applicable).

Logs enable accountability, debugging, and continuous improvement of agent performance.