Skip to main content

Agent Governance & Safety Controls

Status

Owner: Tech lead · Status: draft · Canonical source: flezi.app.ai.agent/docs/governance/agt-policy-overview.md

Describes mechanisms preventing unauthorized autonomous agent actions through AGT Policies and ToolGuard.

ToolGuard Interceptor​

When an AI agent requests a tool invocation, ToolGuard inspects the request prior to execution:

Dynamic Parameter Inspection​

ToolGuard enforces regex and JSON schema validation on tool parameters:

  • Blocks dangerous shell commands (rm -rf, fork bombs).
  • Enforces human approval on write operations to corporate databases.

Existing Controls​

  • Schema validation via agent-governance-toolkit==4.1.0.
  • Mandatory Human-in-the-Loop gating on high-risk operations.

Evidence (Files & Configuration)​

  • flezi.app.ai.agent/app/governance/
  • flezi.app.api.backend/app/features/policies/validator.py

Limitations & Roadmap​

  • Real-time input prompt injection guardrails.