When can an agent close an internal request without a manager?
When can an agent close an internal request without a manager?
Question / origin. In a September 30, 2026 discussion opened from the Claude Code manager-workflow post, Dru distinguished useful summaries from solving request volume. He wants agents to answer and close routine asks, escalating exceptions under explicit principles and boundaries. October 1 clarification by Dru, in discussion of the Taobao post: “I was thinking more about internal AI usage rather than external when I said you want ai to close the loop on it's own.” Read the original proposal as agents resolving internal team/organizational requests, not a claim that customer-facing support should autonomously close tickets. The Taobao study is a warning about handoff design from a different setting, not the primary test population. Dru separately wants early interactive discovery prototypes discarded rather than turning their approach into an anchored implementation plan. The original manager account describes digests and just-in-time planning, not measured agent-owned closure. See planning, roles, gated authority and the prototype test.
Good answer, corrected population. Primary operating record or comparative evidence for internal incoming asks—for example team information requests, routine operational changes, triage or maintenance tasks—with all-incoming denominator, permitted actions, actual requester-confirmed result, exceptions, reopens/rollbacks, active human minutes and the owner of a wrong decision. Distinguish an internal task marked Done from its requester’s verified outcome; distinguish agent completion from a human approval or follow-up gate. Do not silently substitute external customer-support closure for internal software/team work.
Subquestions and evidence. (1) Internal task and action rights: identify which request categories the agent may answer, edit, schedule or repair, what it must escalate and who can reverse it. A public internal maintenance record shows an investigation agent contesting an administrative ‘not planned’ close after observed recurrence; that is a status-integrity example, not a delivered fix. (2) The nearest software-workflow examples remain incomplete: Linear’s first-person account marks an issue Done on merge but reopens an external Intercom request for human CX follow-up; its duplicate-canonical change preserves downstream notification. Klaviyo’s vendor case describes a customer issue → PM delegation → Cursor PR → engineer review → shipped fix in three business days, without source ticket, user confirmation, reopen or human-effort accounting. Neither is a measured internal requester-to-independent-agent-closure chain. (3) External support as bounded contrast, not direct validation: Taobao randomized an external support deployment with same-issue seven-day retrial and ratings; emotionally escalated cases fared worse than matched human-only cases, but escalation types were not randomized. Nubank compares support-agent variants, not human-only handling. Fin separates handoff from claimed resolution and sometimes counts silence as assumed resolution. Bitmovin’s support-side engineer reports guarded direct Zendesk closure without an outcome denominator. These illustrate how premature terminal states or late escalation might hurt, not whether internal autonomous closure works. (4) Prototype anchoring belongs to the separate research test; the visual-ideation experiment is indirect.
October 1 source search and gap. A four-question primary-source search found no single requester-confirmed agent-coded software fix; it largely tested external customer-facing chains and should not be mistaken for a search exhausted for internal requests. Search next for original internal request logs with agent action, requester acceptance, reopens and human-time accounting.
Short primary-source list. Wang et al., randomized Taobao field paper; Gupta et al., Nubank KDD 2026 industry paper; Intercom billing definitions, metric migration, misread reply warning; Linear’s operating account; Linear’s Klaviyo vendor case; October 1 source list in the trace note. Vendor success metrics and controlled experiments answer different questions.
Next test. Follow a cohort of internal team requests: prespecify a low-risk bucket, agent authority and disallowed commitments, exception trigger, owner of mistakes, requester acceptance criterion and independent audit. Measure intake, truly resolved requests, reopens/rollbacks, active human time and harm from wrong closure. The Taobao result can motivate early escalation thresholds, but cannot estimate this internal effect.