Overview

Aegis Sovereign ships four pre-built AI governance agents that run as persistent workers in your Kubernetes cluster. Each agent subscribes to platform events via the internal event bus and acts autonomously — no human trigger required. Agents authenticate with a PAT scoped to the ops role.

Agent	Trigger	Primary Action
CompliancePipelineAgent	`model.registered` event	Runs full eval suite — frameworks, bias, adversarial — gates promotion
IncidentResponseAgent	Drift alert / eval failure / blocked promotion	Posts Slack alert, creates JIRA ticket, pages PagerDuty for critical severity
GovernanceReportingAgent	Scheduled (weekly) or on-demand	Generates executive compliance summary across all production models
RemediationSuggestionAgent	Eval failure with score < threshold	Produces prioritised fix plan — retraining recommendations, dataset gaps, policy changes

CompliancePipelineAgent

Fires on every model.registered event. Runs the following sequence in parallel: (1) evaluate against all active regulatory frameworks, (2) run the AI Safety Gate — red-team certification run (18+ adversarial probes) + ModelSafetyRegistry clearance check, (3) run bias evaluation — Disparate Impact Ratio across all configured demographic groups, (4) run adversarial robustness suite (FGSM, PGD, Carlini-Wagner, DeepFool, Square Attack), (5) populate the model card, (6) write audit chain entry, (7) evaluate all promotion gates. If any regulatory evaluation OR the safety gate fails, the model is set to blocked and an approval request is created with a dedicated Safety Gate Failures section. If all pass, the model is auto-promoted to staging.

bash

Manually trigger compliance pipeline

curl -X POST https://sovereign.yourcompany.com/api/v1/agents/compliance-pipeline/trigger \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "model_id": "mdl_fraud_v4", "force_rerun": true }'

IncidentResponseAgent

Triggered by seven event types — drift alerts (PSI ≥ 0.10), evaluation failures, blocked promotions, and three new safety events: safety_violation.critical, safety_drift.detected, safety_slo.breached. Severity is determined automatically: WARNING (PSI 0.10–0.24, score 60–74), CRITICAL (PSI ≥ 0.25, robustness < 0.40, DIR < 0.60, or any safety_violation.critical). For safety events, the agent also auto-creates an approval request in addition to dispatching to the configured integrations (SIEM via Splunk HEC / AWS CloudWatch; incident notifications to PagerDuty, Slack, and Jira). The env vars AGENT_SLACK_WEBHOOK_URL, AGENT_JIRA_*, and AGENT_PAGERDUTY_KEY configure the notification targets.

bash

Configure incident notification targets

1curl -X PUT https://sovereign.yourcompany.com/api/v1/workspaces/WS_ID/incident-config \
2  -H "Authorization: Bearer $TOKEN" \
3  -H "Content-Type: application/json" \
4  -d '{
5    "slack_channel": "#ai-governance-alerts",
6    "jira_project": "AIOPS",
7    "pagerduty_integration_key": "pd_key_xxxx",
8    "critical_only_pagerduty": true
9  }'

RemediationSuggestionAgent

When a compliance eval, robustness check, or AI safety gate fails, the RemediationSuggestionAgent analyses the failure evidence and produces a ranked remediation plan. Plans are attached to the JIRA ticket created by the IncidentResponseAgent and surfaced in the platform UI under Models → [version] → Remediation Plan. For safety gate failures, the plan includes specific guidance on which red-team probe category failed, which certification test case to fix, and what is required to earn ModelSafetyRegistry clearance.

Failure Type	Typical Remediation Suggestions
EU AI Act score < 75	Add Article 9 risk management documentation; classify use case tier
DIR < 0.80 (bias)	Examine training data imbalance for flagged groups; consider re-sampling or re-weighting
Adversarial robustness < 0.60	Apply adversarial training (FGSM + PGD); add input validation layer
PSI drift > 0.10	Investigate feature distribution shift in input pipeline; schedule retraining