Internal team request → agent action → requester closure: field-audit brief

October 26, 2026. Question. Has an internal organizational or software-team request been independently resolved by an agent, under delegated authority, with evidence the original requester accepted the action and remained served? A good answer uses a dated, same-request log for intake, authority and exception rule, agent action, independent verification, requester response, any reopen and active human attention. Unlike customer support, this tests Dru’s September 30 proposal clarified October 1 as internal loop closure.

Before searching. The internal-closure framing lacked a full intake denominator; Linear stops at merge and CX follow-up; ObjectStack shows a real delegated maintenance action and scheduled result but not the requester’s later quality check or active review minutes. gh-aw shows an agent contesting an administrative closure after recurrence, not a released fix. Incident memory lacks observed second-use retrieval.

Three search questions and primary sources. (1) Has an internal helpdesk request been agent-closed with requester acceptance? The new result is ServiceNow’s implementer playbook and its internal IT case study: bounded low-priority L1 eligibility and resolved rates, no original-employee verification record. Victor Chen’s August 26 operating explainer says sending resolution steps can mark a ticket closed; direct action is a different mode. (2) Does one accepted correction reach the next internal task? Public GitHub searches for reviewer disposition and later task retrieval mostly returned proposals, generic review automation and an older review-skill development screen; no paired internal task and employee acceptance. Our diagnostic synthesis remains the record of separated links; gh-aw has recurring alerts but not adopted repair. (3) What is measured after autonomous closure? The playbook recommends intake coverage, durable auto-resolution without reopen, human intervention, action versus advice, requester waiting and backout; it provides no actual post-close cohort with reopens and human touch minutes. The vendor case study gives 15% of L1 incidents eligible and resolution of 46% of that eligible slice. Separately, the September 15 executive blog and September 29 vendor announcement claim almost 90% of employee IT requests handled and 99% of those interactions resolved; they do not reconcile their population or definition with the L1 cohort. The case study’s '99% faster' is about speed, not the share resolved. A cross-system percentage trend would be invented.

What the read changed. This is no longer a search with only external-customer measures: an internal deployment publishes two explicit cohort fractions and a practitioner spells out the missing outcome ledger. But vendor case material measures a status, not a verified employee repair. Sending advice, making a change, marking resolved, surviving a reopen window and the requester successfully working again are different endpoints. A reopened incident is a useful falsifier, not a sufficient positive oracle: silence and a duplicate ticket can mask an unsolved request. In these sources there is no independently audited action or all-incoming agent-versus-human time comparison. Platform metadata and selected company examples cannot substitute for a randomized trial.

Next discriminating test. For one bounded internal request category, report all incoming, eligible, attempted, advice-only, actual change, refused/escalated, agent-resolved, employee-confirmed at a fixed follow-up and reopened/duplicated within that window, then human touch and model/runtime spend per confirmed resolution. Attach a reviewed decision and recheck artifact to the next case. Stop querying generic 'autonomous desk' headlines; seek original employee confirmation, a timestamped second incident applying a correction, or a released cohort table rather than another aspirational announcement. Link to role shifts, paired acceptance and cost.