Post-Market Monitoring and Incident Reporting¶

Purpose: Define post-deployment monitoring capabilities for AI agent systems and incident reporting procedures as required by EU AI Act Articles 72 and 62.

Deadline: EU AI Act provisions apply from August 2, 2026.

1. Post-Market Monitoring Plan¶

1.1 Overview¶

EU AI Act Article 72 requires providers of high-risk AI systems to establish a post-market monitoring system that actively and systematically collects, documents, and analyzes relevant data throughout the lifetime of the AI system.

AGT provides the technical infrastructure for this through:

Requirement	AGT Capability	Module
Performance monitoring	OTel spans + metrics	`enable_otel()`, `trace_policy_evaluation()`
Decision logging	Tamper-evident audit trail	`AuditLog`, `AuditChain`
Anomaly detection	Policy denial rate monitoring	`agt.policy.denials` metric
Bias monitoring	Advisory classifiers	`CallbackAdvisory` with fairness model
Incident detection	Kill switch, circuit breaker	`agent_os.circuit_breaker`
Trend analysis	OTel metric aggregation	Grafana/Azure Monitor dashboards

1.2 Monitoring Dimensions¶

Dimension	What to Monitor	AGT Data Source	Alert Threshold
Policy effectiveness	Denial rate per rule	`agt.policy.denials` counter	Sudden spike (> 2x baseline)
Trust degradation	Agent trust scores over time	`agt.trust.score` gauge	Score drops below tier threshold
Approval patterns	Approval/rejection ratios	`agt.approval.requests` counter	Rejection rate > 50%
Latency	Policy evaluation latency	`agt.policy.latency_ms` histogram	P99 > 5ms
Error rate	Governance errors/exceptions	OTel error spans	Any non-zero count
Data sensitivity	Session attribute ratchets	`SessionState` values	Restricted-level reached
Advisory triggers	Classifier block/flag count	`advisory_check` audit entries	Increasing trend

1.3 Recommended Dashboard¶

┌──────────────────────────────────────────────────────────┐
│ AGT Post-Market Monitoring Dashboard                      │
├──────────────────┬──────────────────┬────────────────────┤
│ Policy Denials   │ Trust Scores     │ Approval Rate      │
│ (last 24h)       │ (distribution)   │ (last 7d)          │
│ ▃▅▇▅▃▂▁▂▃▄      │ ░░▓▓▓▓▓░░░       │ 78% approved       │
├──────────────────┼──────────────────┼────────────────────┤
│ Eval Latency P99 │ Advisory Blocks  │ Active Agents      │
│ 0.08ms           │ 3 (last 24h)     │ 147                │
├──────────────────┴──────────────────┴────────────────────┤
│ Recent Incidents: 0 P0, 0 P1, 2 P2, 5 P3 (last 30d)    │
└──────────────────────────────────────────────────────────┘

1.4 PromQL Queries for Monitoring¶

# Denial rate by rule (anomaly detection)
sum(rate(agt_policy_denials_total[1h])) by (agt_policy_rule)

# Trust score distribution
histogram_quantile(0.5, agt_trust_score_bucket)

# Approval rejection rate
sum(rate(agt_approval_requests_total{agt_approval_outcome="rejected"}[24h]))
/ sum(rate(agt_approval_requests_total[24h]))

# Policy evaluation latency P99
histogram_quantile(0.99, rate(agt_policy_latency_ms_bucket[5m]))

# Advisory classifier trigger rate
sum(rate(advisory_check_total{outcome!="allow"}[24h]))

2. Serious Incident Reporting (Article 62)¶

2.1 What Constitutes a Serious Incident¶

Per EU AI Act Article 3(49), a serious incident is an incident or malfunctioning of a high-risk AI system that directly or indirectly leads to:

Death of a person or serious damage to health
Serious and irreversible disruption of critical infrastructure management
Breach of obligations under EU law intended to protect fundamental rights
Serious damage to property or the environment

2.2 Reporting Timeline¶

Event	Deadline
Serious incident discovered	Document immediately
Initial report to market surveillance authority	Within 15 days of becoming aware
Updated report (if investigation ongoing)	As new information becomes available
Final report	When investigation concludes

2.3 Report Content (Article 62(2))¶

Field	Description	AGT Source
AI system identification	Name, version, DID	`AgentIdentity`, deployment records
Provider/deployer information	Contact details	Organizational records
Date and circumstances	When and how the incident occurred	`AuditLog.query(start_time=..., end_time=...)`
Description of non-compliance	What went wrong	`PolicyDecision` chain, incident investigation
Corrective actions taken	What was done to remediate	Post-mortem actions
Impact assessment	Who was affected and how	FRIA update

2.4 AGT Evidence for Incident Reports¶

from agentmesh.governance import AuditLog
from datetime import datetime, timezone, timedelta

audit = AuditLog()  # existing audit log

# Pull all events around the incident window
incident_start = datetime(2026, 7, 15, 10, 0, tzinfo=timezone.utc)
incident_end = incident_start + timedelta(hours=2)

events = audit.query(
    agent_did="did:agentmesh:affected-agent",
    start_time=incident_start,
    end_time=incident_end,
)

# Export for regulatory report
for event in events:
    print(f"{event.timestamp} | {event.event_type} | {event.action} | {event.outcome}")
    print(f"  Data: {event.data}")

3. Implementation Checklist¶

Before Deployment¶

[ ] Post-market monitoring plan documented
[ ] OTel observability enabled (enable_otel())
[ ] Alert thresholds configured
[ ] Incident response workflow linked (incident-response-workflow.md)
[ ] FRIA completed (fria-template.md)

During Operation¶

[ ] Dashboard reviewed weekly (or per monitoring plan)
[ ] Denial rate trends analyzed monthly
[ ] Trust score distributions reviewed quarterly
[ ] Advisory classifier effectiveness assessed quarterly

On Incident¶

[ ] Follow incident response workflow
[ ] Preserve audit logs (hash-chained, tamper-evident)
[ ] Assess if serious incident per Article 62 criteria
[ ] Report to market surveillance authority within 15 days if applicable
[ ] Update FRIA if risks have changed

Related: FRIA Template · Incident Response Workflow · Record Retention Policy · EU AI Act Checklist · Tutorial 40 — OTel Observability