AgentGraph: Trace-to-Graph Platform for Interactive Analysis and Robustness Testing in Agentic AI Systems
Zekun Wu, Adriano Koshiyama, Emre Kazim, Sahan Bulathwela, Maria Perez-Ortiz, Seonglae Cho, CRISTIAN ENRIQUE MUNOZ VILLALOBOS, Theo King, Umar Mohammed
Proceedings of the AAAI Conference on Artificial Intelligence
Problems Identified (4)
Agentic execution interpretability: Multi-step agentic AI execution patterns are difficult to interpret.
Manual trace reconstruction: Existing observability platforms require manual inspection of traces to reconstruct agent structure and reasoning.
Agentic execution interpretability: Multi-step agentic AI execution patterns are difficult to interpret.
Manual trace reconstruction: Existing observability platforms require manual inspection of traces to reconstruct agent structure and reasoning.
Proposed Solutions (5)
Trace-to-graph agent observability: AgentGraph converts execution logs into interactive knowledge graphs with agents, tasks, tools, data, humans, and typed relations.
Trace-linked graph provenance: Each graph element links back to its exact trace span to support verifiability.
Graph-based agent robustness analysis: AgentGraph supports qualitative failure detection and optimization recommendations plus quantitative robustness evaluation using perturbation testing and causal attribution.
Trace-to-graph agent observability: AgentGraph converts execution logs into interactive knowledge graphs with agents, tasks, tools, data, humans, and typed relations.
Trace-linked graph provenance: Each graph element links back to its exact trace span to support verifiability.
Results (3)
Verifiable graph elements:
Failure and optimization analysis enabled:
Robustness evaluation enabled:
Research Domain
Explainable/observable agentic AI systems