Show your work
The emerging standard for agent explainability does not ask an agent to explain its reasoning. It asks it to keep receipts — and that is a far more tractable engineering problem.
Provenance, citation grounding, retrieval and prompt perturbation testing, tool-call lineage, memory provenance, workflow replay. Six requirements, and not one of them is about interpretability.
Records, not reasons
Where did this fact come from. Which tool call produced it. What was in memory at the time. Can the whole run be played back. Those are logging questions, answerable by systems built to log, and they do not require anyone to understand a transformer.
That is why this standard will arrive before interpretability does. It asks for something achievable this quarter.
Replay is the one that hurts
Playing back an agent run means every non-deterministic input was captured: model responses, tool outputs, retrieved documents, memory state at each step. Most frameworks capture some of that incidentally and none of it deliberately, which makes retrofitting a rebuild rather than a feature.
Most agents in production today clear none of the six.
Two ends of the same movement
An index now catalogues the technical and safety features of agents actually in deployment, rather than in papers. One document writes down what exists; the other writes down what should be required. Between them a claim becomes a row in a table, and rows are checkable.
Why the timing is not a coincidence
An authority that can demand technical documentation will ask for something adjacent to these six. Explainability is quietly turning into record-keeping — which is unglamorous, unfashionable, and the version of the problem that gets solved.
One of the six is already moving. Memory systems posted their largest gains on temporal and multi-hop questions, and memory provenance is on the list.
UST — From Explainability to Control: The 2026 Executive View of AI Interpretability and Explainability → · arXiv — The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems →