an e-Science 2025 extension of W3C PROV that models AI-agent workflows - prompts, tool calls, model versions, agent interactions - as unified provenance graphs across agent frameworks.
The triangle extended to agents: same Entity-Activity-Agent vocabulary, but resting on trusted infrastructure rather than signatures.
Etymology and backstory
PROV plus AGENT - the W3C vocabulary carried into the agent era by Souza and colleagues in the scientific-workflow provenance community, where tracking what a computation did has been a first-class problem since long before LLMs.
Ecosystem
the scientific-workflow provenance lineage, agent observability and tracing tools, and the survey literature on agent trace provenance.
In codexproof
the Related Work delta in one clause - PROV-AGENT extends PROV to agentic workflows but trusts its infrastructure: no signatures, no cross-domain verification. Same vocabulary, no cryptographic binding, so its graph is exactly as trustworthy as the database that stores it. The paper's move binds the same typed links into content-ids under signatures, so the graph itself survives hostile custody - the difference between a record and evidence.