Fiddler Agentic Observability provides visibility and monitoring for multi-agent AI systems throughout their lifecycle. The platform addresses the complexity of autonomous agent systems by offering hierarchical observability across applications, sessions, agents, traces, and spans. The product enables evaluation and testing of agents during development using curated datasets and experiments with different prompts and responses. It supports stress-testing edge cases before deployment to identify weaknesses early. In production, the platform monitors agent behavior through customizable dashboards and reports. It tracks metrics including hallucination, toxicity, PII/PHI detection, jailbreak attempts, and custom KPIs. The system provides visibility into agent reasoning chains, tool calls, and decision paths across sessions. The platform performs hierarchical root-cause analysis to identify failing spans with full context. It surfaces cross-agent dependencies and bottlenecks that affect system performance. Users can filter, sort, and analyze span attributes to identify critical issues. Fiddler Agentic Observability is powered by the Fiddler Trust Service, which includes Trust Models that run in the customer's environment to score agent and LLM inputs and outputs. These models enable real-time monitoring and guardrails without external API calls. The platform integrates with OpenTelemetry and provides native support for multiple agentic frameworks. It offers a continuous feedback loop between development and production to improve agent performance, accuracy, and safety.