Perspective
Could AI Governance Have Its Own JWT Moment?
Some infrastructure problems get solved once, well, and then simply become the standard everyone builds on. JWT did this for authentication tokens, before it, every team invented its own token format. OpenTelemetry did this for observability, before it, every monitoring vendor had incompatible instrumentation. In both cases, the standard did not win because a committee designed it well in a conference room. It won because enough real deployments converged on the same requirements independently, and someone wrote down what they had in common.
AI agent governance looks like it is at that same point right now. Multiple platforms, built independently, with different priorities and different customers, are converging on a similar shape: intercept a proposed action before it executes, evaluate it against policy, write an audit record. That convergence is itself evidence, not proof, but evidence, that the shape is closer to correct than any single vendor's marketing claims about it.
Here is a candidate for what the required elements of that standard should be, based on what regulatory frameworks already ask for and what production deployments have already found necessary.
Four outputs, from one evaluation
A governance determination, a cost-efficiency metric, a confidentiality determination, and an environmental-resource metric, produced from a single evaluation of a single proposed action, not stitched together from four separate systems afterward.
One record, cryptographically bound to the action
Append-only, written before the routing outcome is finalized, so an audit trail exists regardless of whether the action was permitted or withheld.
Dual exclusion, not just exclusion
When a confidentiality violation is the basis for withholding an action, the flagged content should be excluded from both the audit record and any training data derived from it, while the fact and category of the violation is preserved in both. Exclusion from only one of those two places is a partial solution wearing the clothes of a complete one.
An environmental dimension that includes water, not only energy
Carbon and energy dominate the current sustainability conversation around AI. Independent researchers have already flagged that a carbon-only lens can understate the picture, since energy efficiency and water consumption do not always move together. A conformant record should include both dimensions, not the one that happens to be easier to estimate.
None of this is a claim that one company's implementation should become the standard by default. JWT was standardized through the IETF, an open body, after multiple implementations had already proven the shape out in practice, not before. The realistic path here looks the same: more deployments, more convergence, then a genuine standards process, not a single vendor's specification adopted wholesale.
What is worth naming now, while the shape is still being decided, is what a conformant record should contain, so that whoever eventually writes the standard is working from what production systems have actually needed, not from what sounds complete in a whitepaper.