agentprov

AgentPROV turns agent executions into a source-preserving, evidence-qualified canonical record, and derives W3C PROV from that record. The executions can be captured live or released by other research projects. Every derived PROV statement names its evidence class, the rule that produced it, and the field of the record that justifies it. Corpora are hash-chained from the original raw bytes to the final graph. You can verify and recompute them without re-running any model.

This package is the agentprov command-line program and library. It installs the capture library (agentprov-capture) with it; the draft-only Zenodo uploader (agentprov-zenodo) comes with the zenodo extra.

• Website: https://w3id.org/agentprov/
• User guide: https://w3id.org/agentprov/docs/
• Source code: https://github.com/nickjones39/AgentPROV
• Software on Zenodo: https://doi.org/10.5281/zenodo.23848609
• Corpora on Zenodo (one collection): https://doi.org/10.5281/zenodo.23814758

Install

Python 3.14.8 or later:

pip install agentprov              # or: uv tool install agentprov
pip install "agentprov[zenodo]"    # with the draft-only Zenodo uploader
agentprov info                     # versions, the source registry, extras, host isolation grade
agentprov --help

The desktop app is a separate download for macOS, Windows and Linux, on the releases page. Live runs need the repository checkout: the benchmark backends are separate projects that this package does not contain.

What it does

• Import released agent logs (48 sources in 13 formats, single-agent tool-calling traces) and live captures into one canonical execution record (aer/0.2), through isolated, resource-limited workers.
• Derive semantic annotations (sem/0.1) and W3C PROV-JSON in three bundles: core (observed and derived statements), inference and assessment (agentprov-prov/0.1).
• Validate each stage, verify a corpus's hash chain against its root digest, and recompute the derivation byte for byte without calling a model.
• Package a corpus with its datasheet and checksums for an offline deposit; package never uploads. The separate uploader creates unpublished Zenodo drafts only, and publishing is always done by hand on Zenodo.

The user guide walks through each step on a small demo corpus, and its command reference lists every command's --help.

Citation

@software{jones_agentprov_2026,
  author  = {Jones, Nicholas},
  title   = {AgentPROV: execution provenance for LLM-agent security research},
  version = {0.1.1},
  year    = {2026},
  license = {MIT},
  url     = {https://github.com/nickjones39/AgentPROV}
}

Licence

MIT, © 2026 Nicholas Jones. Third-party code and data licences are listed in THIRD_PARTY_NOTICES.md.