AgentKits

AI Agent Risk Assessment

Describe an agent and see how it classifies. AgentAz™ is a lightweight, design-time vocabulary that documents what an agent is authorized to do — and why — for security review. It pairs with whatever policy engine you run; it doesn't enforce anything itself.

Describe your agent

Tools (allowlist)
Authority (what can it do?)
Controls

Describe an agent or load the example to see its Trust Level, safety breakdown, and a ready-to-use agentaz.json.

Trust Levels at a glance

Trust Level is classified by an agent's worst-case action, not its typical behavior — a high-stakes domain built defensively can still earn a low Trust Level. Every flagship blueprint on AgentKits ships with a full AgentAz specification, so you can review its boundaries before you build. Proven, reviewable architectures stay essential even as models improve at one-shot generation.

  • A1 — Research: read-only, no writes.
  • A2 — Recommend: drafts and scores for review; execution tools absent.
  • A3 — Human-Approved: prepares a payload; a human approves before any mutation.
  • A4 — Limited Autonomy: sandboxed execution with mandatory rollback.
  • A5 — Full Autonomy: end-to-end execution within strict budgets and full audit.