Title: The Hidden Cost of Interruption: Measuring Rediscovery Effort When Coding Agents Assume Control of Paused Tasks

Abstract:

Current benchmarks for coding agents primarily assess their ability to resolve repository issues in a single, continuous session. However, actual software development is far more chaotic; tasks frequently face interruptions, reassignments, reviews, and resumption from partial states established by either a previous agent or human engineer. This study investigates this overlooked aspect through the concept of handoff debt, defined as the rediscovery cost incurred when a predecessor’s contributions are unclear or unfinished.

To analyze this phenomenon, we implemented a takeover protocol that halts a coding agent at predefined handoff points, freezes the repository state, and then evaluates successor agents under four distinct handoff scenarios: repository state alone, raw execution traces, summary notes, and structured notes. Our experimental design involved 75 source tasks, which yielded 181 handoff-point tasks and resulted in 724 takeover runs across various successor models.

The data reveals that context-rich handoffs significantly boost efficiency. Compared to taking over from a repository-only state, handoffs that include context reduced the median number of agent events by 20–59% and decreased cumulative prompt token usage by 42–63%. While improvements in solve rates were more modest and varied by model, the gains in efficiency were consistent across the board. These results indicate that future evaluations of coding agents must go beyond simply tracking whether a task is completed; they must also account for the resource cost required for a subsequent agent to effectively resume the work.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC