What does your AI agent
actually do?

Why a separate spec from model cards?

Agents have a different disclosure surface than models. Tool surfaces, refusal behavior, memory persistence, deployment posture — none of these are captured by HuggingFace-style model cards. Agent Cards is the document that makes those properties machine-readable and comparable.

A canonical example

{
  "agent_card_version": "0.1",
  "agent": {
    "id": "customer-support-tier-1",
    "name": "Tier-1 Customer Support Agent",
    "version": "2.4.1",
    "provider": "Kinetic Gain",
    "description": "First-line customer support. Handles account questions, billing lookups, product orientation. Escalates entitlement changes and refunds."
  },
  "capabilities": {
    "primary_purpose": "Resolve common customer questions and triage incoming support tickets.",
    "models_used": [
      { "name": "claude-sonnet-4-6", "provider": "Anthropic", "role": "executor" },
      { "name": "claude-haiku-4-5", "provider": "Anthropic", "role": "router" }
    ],
    "tools": [
      { "name": "billing-lookup", "side_effects": "read",
        "mcp_tool_card_uri": "https://billing.kineticgain.com/.well-known/mcp-tools/billing-lookup.json" },
      { "name": "ticket-create", "side_effects": "mutating" }
    ],
    "max_context_tokens": 100000,
    "memory_persistence": "session",
    "autonomy_level": "supervised"
  },
  "refusal_taxonomy": [
    { "category": "entitlement_change", "behavior": "escalate_to_human" },
    { "category": "refund_request", "behavior": "escalate_to_human" },
    { "category": "legal_advice", "behavior": "refuse_and_explain" }
  ],
  "deployment": { "environment": "production", "uptime_sla": "99.5%", "regions": ["us-east-1", "eu-west-1"] },
  "safety_posture": {
    "human_in_loop_required": ["entitlement_change", "refund_request", "account_deletion"]
  }
}