Governance / FIELD NOTES
AI agent observability and governance: two questions, one evidence trail
A practical way to connect agent telemetry, ownership and permission decisions without confusing a healthy trace with an authorized action.
Telemetry explains recorded behavior. Governance defines who may act, under which conditions, and who must respond when those conditions fail.
An agent can finish a task quickly, produce a clean trace and still act outside its authority. It can also follow an approved process and fail because a dependency is unavailable. These cases need different responses. Treating them as one generic “agent health” problem makes both investigation and accountability harder.
For this article, observability means using recorded signals to understand system behavior. Governance means the decisions, responsibilities and constraints that determine how an organization permits and oversees that behavior. This is a working distinction, not a claim that vendors or standards use identical boundaries.
OpenTelemetry describes traces as sequences of spans representing operations, with information such as timing, attributes and relationships. Those records are useful building blocks for investigation. They do not, by their presence alone, establish an organization's permission decision. OpenTelemetry's trace documentation explains the underlying model.
Begin with the question a reviewer needs to answer
Consider a hypothetical support agent that updates a customer's delivery address. The engineering team asks whether the call completed, how long it took and which dependency failed. The operations owner asks whether the customer requested the change, whether the agent could modify that order, and whether a human approval was required.
The first set of questions needs execution records. The second needs context that may live outside the trace: an ownership register, a policy version, an approval decision or the original customer instruction. A dashboard cannot manufacture these missing facts by attaching a confidence score to the call.
Start the design with a small list of review questions. For each question, identify the owning source, the person responsible for its interpretation and the conditions under which the answer remains unknown. This makes the collection requirement concrete before anyone adds another event field.
Use a shared reference, not a shared verdict
We recommend linking the task, attempted action, identity, policy decision and destination result with stable references. Keep the source and timestamp of each observation. An action record may say that an API returned success; an approval record may say that a particular address change was allowed. These statements should remain independently inspectable.
Avoid collapsing them into a single “compliant” flag. That flag becomes difficult to explain when the policy changes, the destination reverses the update or an approval turns out to reference a different order. Separate records allow the reviewer to see the disagreement rather than inherit an earlier interpretation.
This separation also helps teams divide responsibility. A platform engineer can repair a broken collector without deciding business policy. An operations owner can tighten an approval requirement without rewriting the historical execution record.
Make missing context a visible state
For the hypothetical address-change workflow, a useful review could distinguish these conditions:
- The agent reported a successful call, and a destination record confirms the matching change.
- The call was recorded, but the destination evidence has not arrived.
- The destination shows a change that has no matching agent record in the reviewed period.
- An approval exists, but its order or proposed address does not match the action.
- The action is attributable to a credential, but its accountable owner is unknown.
These are proposed review states, not a universal taxonomy. Their value is that each points toward a specific next step. “Waiting for destination evidence” should trigger a collection check. “Approval mismatch” should trigger a review of the action and its authority.
Keep governance broader than instrumentation
NIST describes its AI Risk Management Framework as voluntary guidance for incorporating trustworthiness considerations into AI design, development, use and evaluation. Instrumentation can support that work; it does not replace organizational decisions about acceptable use. NIST's framework overview is a useful starting point.
Our recommendation is to make an owner answer three questions before expanding autonomy: what the agent may change, how that boundary is enforced, and what evidence a reviewer will receive after a change. If any answer is missing, record the gap and decide explicitly whether the workflow can proceed.
A practical first implementation
Choose one consequential workflow rather than attempting to classify every agent at once. Document its allowed actions and destination. Instrument one complete journey, including a rejection, a timeout and a later correction. Ask a colleague who did not build it to reconstruct what happened from the exported records.
If that reviewer cannot identify the action, its owner, its authority and the observed result, improve the missing connection. More telemetry is helpful only when it answers a question that matters. The goal is a reviewable account of a bounded workflow, with uncertainty preserved wherever the available evidence cannot support a stronger conclusion.
Sources and further reading
Primary sources checked on 16 November 2026. Examples in this article are illustrative unless stated otherwise.