Agent Guardrails — Tiered HITL Governance for AI Agent Writes

The problem

Autonomous agents are safe when reading data but represent a massive liability when granted write access. Without strict governance, agents can trigger irreversible side effects, drain funds, or corrupt production databases through bulk operations.

What it does

• Classifies every proposed agent action into one of four risk tiers (Deny, Tier 3 Human, Tier 2 Confirm, Tier 1 Auto).
• Enforces an non-negotiable human-in-the-loop (HITL) requirement for high-stakes domains like payments, auth, and PII.
• Prevents destructive "Deny" actions (like truncate or mass delete) from ever reaching an approval stage.
• Generates a mandatory, append-only JSON-lines audit trail for every governance decision made.
• Provides a reference Python logic for integrating these gates into existing agentic loops.

Frameworks & tools

Technology-agnostic logic. Reference implementation available in Python with zero third-party dependencies.

Why this beats prompting it yourself

LLMs are notoriously bad at self-policing and often hallucinate that a dangerous action is "safe." This skill provides a deterministic classification procedure that favors over-asking over catastrophic failure, ensuring the agent's self-assessment never overrides hard-coded safety invariants.

Use cases

• Governing financial agents that execute refunds or payouts over specific thresholds.
• Adding an approval gate to DevOps agents before they deploy to production or modify infrastructure.
• Protecting CRM data from bulk deletion or unauthorized PII exports by an AI.
• Managing database migrations where irreversible mutations require manual sign-off.

Known limitations

Requires an implementation of an "Approver" interface (like a CLI prompt or Slack webhook) to handle Tier 2 and Tier 3 actions.