Agents

An agent is the unit you protect. It can be a chatbot, a tool-using LLM in your backend, an autonomous worker, or a human operator running scripts. Whatever it is, you give it an identity and Detecte tracks every action it attempts.

Properties

Field	Description
`id`	Stable identifier (`agent_…`). Generated.
`name`	Human-friendly name (`support_bot`, `treasury-rebalancer`).
`description`	What the agent does. Surfaced in the dashboard.
`tier`	`low` \| `medium` \| `high` \| `restricted`. Set by the KYA battery; can be overridden manually.
`declared_capabilities`	Tools the agent is allowed to use. Used for policy matching.
`risk_score`	A rolling 0–100 score Detecte computes from recent decisions and anomalies.

Registering an agent

Auto-registration happens the first time you verify() with a new agent name in test mode. In production, register explicitly:

const agent = await detecte.agents.create({
  name: "support_bot",
  description: "Customer-facing refund and account agent",
  declared_capabilities: ["refund_order", "lookup_account"],
});

Tiers and the KYA verification

Tiers come from the KYA (Know Your Agent) test battery. Run it once when you ship a new prompt or new toolset:

detecte scan agent_support_bot

The battery probes the agent for prompt-injection resistance, jailbreak susceptibility, capability boundary respect, and seven other categories. Output is a 0–100 score, a tier, and a signed JWT credential valid for 30 days.

When the credential expires, Detecte will automatically downgrade the agent's tier — you re-run scan after each major prompt change.