Marketing Hub
Local Demo

Agent Observability

Collapse / Expand

Agent Observability is the operational workspace for agent observability within assistant and agent operations, where teams manage concrete records, dependencies, and execution decisions rather than static summaries. The business and compliance goal is to keep logs, evals, hallucination tests aligned with approved claims, consent controls, market rules, and audit evidence at every step. This page presents role-scoped actions, linked downstream impacts, and policy-aware guardrails so users can work quickly without violating promotional or safety obligations. Next, use this agent observability workspace to execute governed tasks and route outcomes to connected teams.

RestrictedSection: Assistant + AgentsSeed: 20260224

Access Boundaries

SSO + RBAC/ABAC controls are evaluated by brand, market, function, agency zone, and risk tier before actions are enabled.

Field Surfaces: Approved-only enforcedExternal Partner: Submit-onlyRisk Tier Scope: Low/Moderate/High/Critical
Agent Observability: Agent Runs
182-6% governed
Agent Observability: Policy Pass
189+7% governed
Agent Observability: Human Gate Time
196+3% governed
Agent Observability: Traceability Score
203-1% governed
Agent Observability: Output Quality
210-5% governed
Agent Observability: Eval Pass
217+8% governed

Operational Focus

  • Inspect run latency and failure classes.
  • Track hallucination test pass rates.
  • Compare model behavior across roles.

What To Do Next

  • Open failing test trace.
  • Adjust policy guardrail.
  • Approve model registry update.

Agent Observability Studio

Monitor agent logs, evaluation outcomes, hallucination tests, and model performance profiles.

IDTitleMarketOwnerStatusRiskRelated
AG-0001Agent Observability · Agent execution 1MXAgent OpsBlockedHighAgent ObservabilityAssistant
AG-0002Agent Observability · Policy check 2SGHuman ReviewerEscalatedCriticalAgent ObservabilityAssistant
AG-0003Agent Observability · Human approval 3USPolicy AdminDraftLowAgent ObservabilityAssistant
AG-0004Agent Observability · Output routing 4CAAgent OpsIn ReviewModerateAgent ObservabilityAssistant
AG-0005Agent Observability · Agent execution 5UKHuman ReviewerApprovedHighAgent ObservabilityAssistant
AG-0006Agent Observability · Policy check 6DEPolicy AdminActiveCriticalAgent ObservabilityAssistant
AG-0007Agent Observability · Human approval 7FRAgent OpsBlockedLowAgent ObservabilityAssistant
AG-0008Agent Observability · Output routing 8ITHuman ReviewerEscalatedModerateAgent ObservabilityAssistant
AG-0009Agent Observability · Agent execution 9ESPolicy AdminDraftHighAgent ObservabilityAssistant
AG-0010Agent Observability · Policy check 10JPAgent OpsIn ReviewCriticalAgent ObservabilityAssistant
AG-0011Agent Observability · Human approval 11AUHuman ReviewerApprovedLowAgent ObservabilityAssistant
AG-0012Agent Observability · Output routing 12BRPolicy AdminActiveModerateAgent ObservabilityAssistant

Relationship Graph

Linked entities are enabled so users can traverse parent-child dependencies and downstream impacts without manual lookup.

Command CenterSearch

Agent Hub Tools

Human Gate

Capture approval/reject decision and route output.

MLRAudit
Execution StepOwnerETA
Validate inputOps5m
Policy checksCompliance8m
Route next actionWorkflow3m

Governed Assistant Output

Grounding sources include approved claims, PI/SmPC requirements, references, strategy docs, and previously approved assets. Responses include citations, confidence, output classification, request-MLR CTA, and prompt-output audit logs.

Classification: PromotionalConfidence: 0.91

Agent Hub

Model-agnostic registry, tool permissions, policy checks, and human-in-the-loop approvals are deterministic for local demo runs.

Model Registry

  • mdl-gpt-enterprise · OpenAI · high
  • mdl-claude-enterprise · Anthropic · high
  • mdl-local-llama · Self-hosted · sandbox

Sample Run

1. Input Loaded
2. Policy Guardrails Checked
3. Source Retrieval + Traceability
4. Human Approval Gate
5. Final Routed Artifact

Risk Flags: Missing fair balance in social template (resolved), off-label drift score 0.03 (pass), unapproved claim attempt blocked.

Human Approval Phase: Medical reviewer sign-off required before output release.

Final Artifact: Routed to `/mlr/queue` with note `Awaiting final promotional approval`.

Audit Entry: EVT-AI-2026-02-24-04122 recorded with tool permissions snapshot.

Output Classification: Promotional · Confidence: 0.91
  • AST-00431: HCP Efficacy Leave-behind v3
  • CLM-0214: Primary endpoint response claim
  • REF-0550: SmPC Section 4.8 safety reference

Agent Catalog

Market Landscape AgentBrand Positioning AgentCampaign Brief AgentSegmentation AgentClaims Librarian AgentClaim-to-Asset Traceability AgentFair Balance / Safety Presence AgentOPDP/Social Space Constraint Checklist AgentMLR Pre-Check AgentModular Asset Composer AgentVariant Generator AgentLocalization Brief AgentSEO/Discoverability AgentNext Best Action AgentJourney Builder AgentField-Digital Coordination AgentPerformance Analyst AgentContent ROI AgentExperiment Design AgentAttribution Insights AgentWorkflow Triage AgentAsset Reuse Finder AgentExpiry & Withdrawal AgentAgency QA AgentAdverse Event Detector AgentProduct Complaint Classifier Agent

Governance Notes

  • Every approval, publication, policy block, and AI output writes an audit event.
  • Consent and frequency caps are evaluated before deployment packaging.
  • Expiry and withdrawal controls maintain cross-channel evidence status.
  • Safety routing is always flag-and-route with human adjudication.

Geo Coverage

US1,820 active
UK640 active
DE550 active
JP730 active
BR500 active
AU320 active

Seed Coverage Summary

brands: 8campaigns: 30briefs: 200assets: 6,000components: 20,000claims: 2,500references: 8,000markets: 12localizationJobs: 1,500mlrPackages: 15,000reviewSteps: 60,000annotations: 200,000interactionsRollup: 2,000,000socialSignals: 50,000aeSignals: 1,500experiments: 5,000reports: 3,000connectors: 250auditLog: 300,000