Trust Scores and Agent Drift Detection

Estimated reading time: 3 minutes

Trust scoring and drift detection answer different questions. A trust score is a point-in-time measure of how much standing an agent has earned; a drift alert signals that an agent's current operating behavior has moved away from its own established baseline. Both are under Monitor in the workspace, and neither is behind a separate feature flag.

Trust score

Every agent has a trust score from 0–100, computed from five weighted components: identity verification, behavior history, guardrail compliance, reputation, and security posture. The score maps to one of five trust levels — untrusted, limited, standard, verified, trusted. The weighting and tier thresholds are internal, tuned parameters and not published here; what follows is the shape of the API surface.

Third-party attestations

An external party — a certification body, auditor, or your own security team — can contribute a signed attestation that adjusts an agent's reputation component. Submission requires:

  • A registered provider. Only a platform administrator can register an attestation provider's signer key; submissions from an unregistered provider are rejected.
  • A verified signature. Every submission carries an ECDSA-P256-SHA256 signature over the asserted fields, checked against the registered signer key before the attestation is accepted.
  • The organization owner role, to submit or revoke an attestation.

An attestation's contribution to the score is bounded — it cannot move an agent across trust-level tiers on its own. Every submission and revocation writes a signed audit-log entry, and submission is rate-limited.

API

Method & path What it does
GET /organizations/{orgId}/agents/{agentId}/trust/score Current trust score, level, and component breakdown
GET /organizations/{orgId}/agents/{agentId}/trust/history?limit= Score history (defaults to 90 entries)
GET /organizations/{orgId}/agents/{agentId}/trust/attestations List attestations for the agent
POST /organizations/{orgId}/agents/{agentId}/trust/attestations Submit a signed attestation (organization owner only)
POST /organizations/{orgId}/agents/{agentId}/trust/attestations/{attestationId}/revoke Revoke an attestation (organization owner only)
GET /organizations/{orgId}/trust/agents-summary Trust scores for every agent in the organization in one response

Drift detection

Praesidia builds a behavioral and operational baseline per agent and raises a drift alert when current activity deviates from it. An alert carries a drift type — response time, error rate, cost, token usage, tool-usage distribution, output length, or behavior pattern — a severity (low, medium, high, critical), the baseline and current values, and the deviation percentage. An alert's status moves through open, acknowledged, resolved, or dismissed as your team works it.

A related, finer-grained signal scores individual calls against the agent's learned profile: known tools, its typical position in a chain, and the per-tool argument key-shapes it normally uses — argument values are never inspected, only which keys are present. An agent with no usable baseline yet is flagged coldStart rather than scored.

API

Method & path What it does
GET /organizations/{orgId}/agents/drift/alerts List drift alerts, filterable by agent, severity, status, and date range
GET /organizations/{orgId}/agents/drift/alerts/{alertId} Get one alert
POST /organizations/{orgId}/agents/drift/alerts/{alertId}/acknowledge Acknowledge an alert
POST /organizations/{orgId}/agents/drift/alerts/{alertId}/resolve Resolve an alert
POST /organizations/{orgId}/agents/drift/alerts/{alertId}/dismiss Dismiss an alert
GET /organizations/{orgId}/agents/drift/{agentId}/health Baseline, current metrics, and active alert counts for one agent
GET /organizations/{orgId}/agents/drift/{agentId}/behavior The agent's learned behavioral profile and recent per-call deviations
POST /organizations/{orgId}/agents/drift/{agentId}/baseline Manually rebuild an agent's baseline

Reading drift alerts and the behavior profile requires the agents.view permission; acknowledging or resolving an alert requires agents.update; rebuilding a baseline requires agents.configure.

In the workspace

Trust scores and attestations are on Monitor → Trust; drift alerts and the behavior panel are on Monitor → Agents → Drift.

Next steps