When can an agent close a 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. He 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. Primary operator record or experiment with incoming denominator; actual resolution, recontact, escalation and human-touch counts; action rights and reversal owner; inspectable request-to-outcome examples. Distinguish a support reply from a verifiably resolved case and approval gate from audit sampling.

Subquestions and evidence. (1) End-to-end closure: Taobao provides a randomized deployment, same-issue seven-day retrial and ratings, but no independently checked customer-level terminal state. Nubank A/B-tested self-service and tNPS against old agents, not human-only. (2) Rights/handoff: Fin separates human-free resolution, intentional configured handoff, escalation and pending; high-risk actions can require human approval. Nubank uses narrow tools and supervised action rights. (3) Wrong closures and audit: Intercom documents silence counted as assumed resolution and ‘willing to wait’ misclassified as confirmed resolution. Its 2026 metric-definition change increases headline resolution rate without changing automation across the incoming queue. Taobao’s emotionally escalated subset had worse seven-day retrials and ratings than matched human-only cases; escalation types were not randomized. (4) Internal software requests: Klaviyo’s vendor case describes one customer issue → PM delegation → Cursor PR → engineer review → shipped fix in three business days, but no source ticket, user confirmation, reopen or human-effort accounting. Linear’s own first-person composite says issue marked Done on merge reopens Intercom request for CX to notify the customer; automation of software completion deliberately creates a human closure step. These two firms’ stories are separate and neither shows agent-owned customer closure. (5) Prototype anchoring belongs to the separate research test; the visual-ideation experiment is indirect.

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 named operating account; Linear’s Klaviyo vendor case. A vendor success metric and a controlled experiment answer different questions.

Still missing / next test. A matched software-team request-to-repair trace with agent answer/action, consumer confirmation, human attention and reopen/rollback. Prospective intervention: prespecify low-risk request bucket, allowed answers/actions and disallowed commitments, escalation thresholds and owner, measure all-incoming request allocation, verified closure, seven-day+ recontact, human active minutes, customer rating and wrong closures. Support-study result is a stress test on the handoff, not a general no-go verdict on agent closure. In the Linear-style workflow, test whether the customer-follow-up stage can be safely delegated under bounded rights, rather than conflating ‘Done’ with closure.