AI wrote the code. Can you prove it followed the rules?
AI agents now write, review and ship code, and our supply-chain security can't tell how that code was made. A green build and a clean scan don't tell you which model or tools the agent used, whether it had your security guidelines, or whether a human actually reviewed the change.
In this talk I introduce Agentic Process Evidence, a proposed open community standard. It records the AI development process as structured, signed evidence that is attached to the commit and linked back to the full agent session logs, so release policies can enforce it automatically as policy as code.
What you'll learn:
✅ The "provenance gap" in AI-assisted software development
✅ Why session logs alone aren't enough for governance
✅ The two-layer model: session logs + process evidence
✅ How to gate releases on approved models, tools, guidelines and human review
✅ How to trace every affected release when an MCP server turns out to be malicious
Note: Provenance doesn't prove that generated code is correct. It tells you what was recorded about the process, and gives you a foundation for verification.
🔗 Spec on GitHub (working draft, contributions welcome): https://github.com/jfrog/agentic-process-evidence
📝 Blog post: https://jfrog.com/blog/governing-the-agentic-sdlc/
Chapters:
0:00 AI wrote the code — can you prove it followed the rules?
0:29 A typical AI-assisted change
0:54 4 questions every release approver should ask
1:57 The provenance gap: a green build isn't enough
2:13 Why session logs alone aren't enough
2:57 Structured, verifiable evidence (policy as code)
3:17 Introducing Agentic Process Evidence
3:27 Two layers: session logs + process evidence
3:55 What the evidence records
4:19 Attached to the commit, signed & verifiable
4:40 Enforcing release policy with evidence
5:36 When an MCP tool is compromised
5:59 Provenance ≠ correctness
6:26 A proposed open community standard
6:58 Wrap-up & how to get involved