AgentAz™ — open-source AI-agent governance + 62 reference blueprints

A small, open spec for writing down what an AI agent is allowed to do — and what its worst-case action is — before you ship it.

AgentAz is design-time only. It doesn't run anything or wrap your code; it's a vocabulary for documenting an agent's authority, failure modes, and escalation path in a file you can review, diff, and hand to a security or compliance team. Your runtime still does the enforcing — AgentAz just makes the intent reviewable.

This repo holds the spec and a reference registry: 62 agent blueprints (13 flagship), each tagged with a Trust Level, a written-out worst-case action, and a machine-readable agentaz.json that validates against the v1.0.0 schema.

The browsable version — search, the risk-assessment tool, downloads — lives at https://www.agent-kits.com. This repo is the raw source.

What's in here

SPEC.md                          The AgentAz v1.0 spec
RATIONALE.md                     Origin & design rationale (the 'why')
CROSSWALK.md                     Mapping to NIST AI RMF, ISO 42001, OWASP Agentic
rfcs/                            Public, dated change process (RFCs)
reference/                       Reference validator — compute a tier from a spec
schema/agentaz-v1.0.schema.json  JSON Schema (draft-07) the contracts validate against
kits/<slug>/agentaz.json         The governance contract for one agent
kits/<slug>/kit.json             Full blueprint: prompt, tools, workflow, examples, failure modes
kits/<slug>/run.py               Runnable demo: native tool-use loop + enforced approval gate
kits/<slug>/evals/run.py         Deterministic safety check for the gate (no API key needed)
kits/<slug>/requirements.txt     Python dependencies for the demo
kits/<slug>/README.md            Plain-English summary + attribution
flagships.json                   The 13 reference kits
LICENSE (Apache-2.0)             Code & schema
docs/CONTENT-LICENSE.md (CC-BY-4.0)  Spec text & blueprint content

Using it

• Validate any agentaz.json against schema/agentaz-v1.0.schema.json (JSON Schema draft-07).
• Enforce it in CI — add the agentaz-validate GitHub Action to any repo; it computes the Trust Level on every push/PR and can fail the check when an agent is riskier than your policy allows.
• Score any agent's governance — paste a system prompt or an agentaz.json into the free, deterministic AgentAz Compliance Scanner; it returns pass/fail gates mapped to Microsoft’s published governance guidance, plus a fix-list.
• Run any blueprint live — each kit page has a browser playground that executes the real loop with your own API key, with the same approval gate as run.py.
• Use the vocabulary in your own agents — it's open, and using it doesn't imply we've reviewed or endorsed your agent.
• Browse the rendered blueprints at https://www.agent-kits.com.

Status

AgentAz v1.0.0 is frozen — no breaking changes within 1.x, and the schema $id URL (https://www.agent-kits.com/agentaz/agentaz-v1.0.schema.json) is permanent. It's early: expect the spec to grow at the edges (more examples, clearer guidance) without breaking what's here. Issues and PRs welcome — see CONTRIBUTING.md.

License & names

Code and schema are Apache-2.0 (LICENSE). Spec text and blueprint content are CC-BY-4.0 (docs/CONTENT-LICENSE.md). “AgentAz” and “AgentKits” are trademarks of AgentKits — the open licenses cover the work, not the names.

How this repo is maintained

It's generated from the main AgentKits source and published one-way, so don't hand-edit files here — changes happen upstream and the repo is regenerated. File issues/PRs per CONTRIBUTING.md.