The Challenge
A result should be checked before it is accepted, and a risky tool should not run until someone allows it.
Agent Pattern
A guardrail checks the output. A policy allows or denies a tool before it runs. Approval waits for a person.
PraisonAI Features
- Guardrails on the result
- Allow and deny rules for tools
- Human approval before a risky tool
- An audit log of tool calls
- The check runs where you host the agent
Checks, not certification
These are checks on an agent’s work, not a compliance certification. Tool calls can go to an audit log today. The record for regulators is part of the handover, which is in early access.