Verification Without Forgetting: Information Prominence and Interaction Distance in Long-Horizon LLM Agent Workflows

Abstract

AI coding agents frequently encounter configuration information early in a multi-step workflow and must later act correctly on the current environment state, which may change silently between workflow steps. This study investigates whether a large language model (LLM) coding agent re-verifies previously encountered information after such a silent mutation, and isolates the process cost of that re-verification from the mechanical cost of completing unrelated intervening tasks. We conducted a controlled 4 (scenario) × 2 (information prominence) × 2 (interaction gap) × 10 (repetition) factorial experiment, yielding 160 independent trials evaluating gemini-3.1-flash-lite. The agent re-verified the active environment configuration in 160 of 160 trials (100%) and achieved a 98.1% task success rate. Critically, we distinguish two process-cost metrics: cumulative workflow tool calls and Phase-D tool calls, isolating the final target-task adaptation after the silent mutation. Under the cumulative metric, interaction gap appeared to strongly affect process cost, but this effect was confounded by the additional intermediate tasks present in long-gap trials. Isolating Phase D removes this confound. Information prominence remained a large, highly significant driver of process cost (β = +2.800 tool calls, p = 2.98 × 10⁻¹³, partial η² = 0.541), while interaction gap alone showed no statistically significant independent effect on Phase-D cost (β = −0.475, p = 0.177, partial η² < 0.001). A significant interaction (β = +1.050, p = 0.036, partial η² = 0.029) indicated that interaction gap compounds the cost of locating non-prominent information. Three task failures, all in the API scenario, occurred after successful re-verification, confirming that environmental adaptation and downstream code correctness are distinct constructs. Under the tested conditions, the evaluated agent reliably re-verified the environment rather than relying on stale information, and the architecture of information placement—not interaction distance alone—was the dominant determinant of verification efficiency.