← Writing

Technical note · 23 July 2026 · Ron Bronson

What Counts as Proof That an Agent Did Its Job?

A record should distinguish a request, an attempt, a provider acknowledgment and a confirmed outcome.

An AI assistant says it sent a message. That statement tells the person what the assistant believes happened. The mail service may have accepted the request, rejected it or queued it for delivery. The recipient may still have received nothing.

A meaningful record has to keep these facts separate.

An action has several states

The Vloora research model distinguishes requests, authorization, attempts, acknowledgments, decisions and contested outcomes. A more detailed transaction log may show when an action was proposed, approved, attempted, accepted by a provider or confirmed by independent evidence.

The exact state names may change as implementation develops. The distinction matters more: a tool call is evidence of an attempt, and a provider-generated receipt supports only the assertion that provider made.

Follow one request

Consider an assistant asked to email a weekly summary. A review record might include:

Recorded eventWhat it establishes
Assistant proposed sending a messageA request exists
Account owner approved the exact content and recipientAuthorization was granted for that action
Controlled adapter attempted a sendSoftware made an execution attempt
Mail provider returned a message IDProvider accepted a request
Provider reported deliveryDelivery was reported by that provider

If the final event never arrives, the honest status is that delivery is unconfirmed. A receipt must name its issuer and retain the evidence on which the displayed status rests.

Records need corrections

External services can change their mind or issue decisions after an initial acknowledgment. A renewal can be submitted and later rejected. A payment can be accepted and later reversed.

A durable history should append those later events rather than replace the original entry with a generic success label. It should be possible to distinguish the assistant's account, the observed request and the receiving organization's final disposition.

What the person should see

The interface can use ordinary language: “You approved this message,” “Accepted by email service,” or “Delivery not confirmed.” Technical details can be retained behind that summary. The consumer should be able to export the record without exporting their credentials.

The same distinction supports performance assessment. Before a product receives credit for a successful task, the record needs to establish which part of the task actually succeeded.

Sources

Research and sources updated 16 October 2026.