Context Engineering for Enterprise Analytical Agents: An Observability-Driven Study of Context Accumulation, Planning Amplification, and Shared-Prefix Cache Behaviour Across Two Governed Production Agents

Enterprise analytical agents accumulate context across multi-turn conversations, but the cost is rarely measured. We report an observational study of two independently configured production analytical agents -- governed text-to-SQL systems over semantic layers -- using platform-native observability telemetry covering 784 threads and 2,269 messages. Our principal finding is that prompt cache residency is not governed by a user's own pause alone. Within a reproducible five-minute own-thread reuse boundary, turns showed near-total reuse (98.1% retention, 895 median uncached tokens). Beyond that boundary, own-thread reuse fell sharply and outcomes differed strongly with recent activity from other threads on the same agent: at gaps over ten minutes, turns with recent cross-thread activity retained cache on 77.3% of occasions against 12.2% without it, with 36% lower median uncached input. Retention generally increased with the agent's request rate on both agents, from 8.7% to 69.6%. The pattern is consistent with partial shared-prefix reuse and motivates architectural hypotheses about traffic consolidation that require controlled validation. Three defensible exposure metrics also diverge on one workload: median first-step input rose with thread depth (68,429 to 208,421 tokens) while declining planning depth (3.49 to 1.60 calls per message) held the per-message envelope flat, and rising cache reuse cut mean uncached input 74%. Each alone misleads. Per-step input was flat, locating growth in conversation depth rather than orchestration depth. Structure replicated across both agents; absolute constants did not, with turn-1 baseline differing 2.4x. We propose a three-factor exposure decomposition and a governed measurement lifecycle, and document two self-corrections, including a bucketing choice that concealed the conditioning variable above.