The Scope Creep Myth: How Hierarchical Visibility Drives Escalation Rather Than Departure in Multi-Agent Systems

When an orchestrator delegates a subtask to a worker agent, it must decide how much context from the supervising hierarchy to expose. Current engineering practice diverges sharply: systems either isolate workers to their immediate assignment or forward the full execution transcript, while literature on over-eager agents suggests that broader visibility invites scope creep. This preprint evaluates that assumption through a pre-registered, controlled experiment. Across 15,589 task delegations on two frontier reasoning models (DeepSeek V4 Flash 0731 and Meta Muse Spark 1.3), a scripted orchestrator delegates assignments across five task domains, each embedding an intentional, benign conflict with the principal's overarching objective. A six-rung visibility ladder with length-matched and hierarchy-matched controls varies what the worker sees, while deterministic, audited checkers score literal adherence, explicit escalation, and unprompted departures toward the root objective. Objective-directed departures occur rarely, remain confined to a single model family, and are almost universally announced within the worker's structured report. Supervisory visibility predominantly induces upward escalation rather than unauthorized action: items with flagging responses rise from 12 to 89 on Muse Spark and from 29 to 88 on DeepSeek between isolation and full ledger visibility. This behavior depends on semantic coherence rather than context length: hierarchy borrowed from an unrelated task doubles reasoning token volume without producing departures, while causing Muse Spark to refuse in one-third of runs. A live deployment across 600 multi-agent task runs incorporating real file systems, concurrent sibling agents, peer polling, and an escalation channel yielded zero attributed departures on either model (exact binomial 95% upper bound: 1.24%). The live suite demonstrates that write collisions occur in every concurrent run and that standard un-attributed diffs misclassify inherited sibling updates as scope creep. Exposing the hierarchy approximately doubles reasoning tokens, latency, and inference expenditure while functioning primarily as a mechanism for upward exception reporting. The paper concludes with an architectural audit of 62 orchestration frameworks documenting default visibility mechanisms, alongside data-grounded system design principles. This record contains the preprint PDF and its complete LaTeX source package. Pre-registration documents, stimulus items, deterministic checkers, execution traces with gateway billing logs, audit annotations, analysis scripts, and the Suite B live harness are deposited separately in the companion data and code record.