Reproduce “PROV-AGENT: Unified Provenance for Tracking AI Agent Interactions in Agentic Workflows” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Renan Souza, Amal Gueroudji, Stephen DeWitt, Daniel Rosendo, Tirthankar Ghosal, Robert Ross, Prasanna Balaprakash, Rafael Ferreira da Silva

PROV-AGENT extends workflow provenance to connect agent tools, prompts, model invocations, responses, and decisions with workflow data and tasks. Implemented in Flowcept, it records these links so users can query the inputs and later activities associated with a decision. The authors demonstrate five query examples on a distributed workflow mimicking additive manufacturing, not a live deployment or error detector.

Key insights

  • Unifying AI agent-specific metadata (prompts, responses, model invocations) with traditional workflow provenance under a single W3C PROV-extended model enables end-to-end traceability for complex agentic systems.
  • Real-time instrumentation of AI agent frameworks (like MCP) allows for comprehensive capture of dynamic decision-making processes, which is crucial for understanding and debugging non-deterministic agent behaviors.
  • Detailed provenance graphs can provide the necessary context to perform root cause analysis for issues like AI hallucinations, understand error propagation, and support continuous improvement of AI agents in iterative workflows.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.