Three silent cost leaks: subsumed retrieval, similar scripts, test re-execution
Sep 25 Hu et al.: 1,200 trajectories — three behaviors hit 79–98% of tasks and up to 22.75% of cost; structure-aware retrieval can raise spend 28%; developer-designed skills cut cost up to 41.73%.
Source: Analyzing and Mitigating Cost-Inefficient Behaviors in Coding Agents
Hu, Jiang, Liang, Dey, Wu, and Tan’s Sep 25 study is a receipt book for the verification tax continuous factories pay in tokens, not just in false greens. They analyze 1,200 trajectories from Claude Code and Mini-SWE-Agent across four configurations on SWE-bench Verified, then evaluate mitigations over 10k+ trajectories on held-out Verified and Pro tasks.
Three recurring behaviors cover 79–98% of tasks and account for up to 22.75% of task monetary cost: subsumed retrieval (re-reading code already covered by a prior read — up to 92.33% of tasks / 11.41% of cost), similar script generation (regenerating near-duplicate probes and patch scripts instead of editing persisted artifacts), and test re-execution (rerunning identical tests without an updated patch). Architecture matters: Claude Code’s cross-agent subsumption comes from subagents returning summaries the main agent then re-reads; Mini-SWE-Agent pays more for zoom-in localization, low-feedback shell edits, and long debugging loops. Structure-aware retrieval (CodeGraph) is not a free win — verbose feedback and altered delegation raise end-to-end cost by up to 28.14% in some cells. Agent-synthesized skills from traces give at most ~22% savings; developer-designed, high-level, trace-agnostic skills cut cost by up to 41.73% while shrinking all three behaviors.
Why it matters for code review
Always-on agent fleets compound waste the same way they compound merges. Review overnight spend the way Hu et al. attribute cost: ask which steps were subsumed re-reads, regenerated scripts, or no-op retests, and whether the harness ships reusable artifacts and fail-closed test reuse rather than hoping the model invents frugality. Steady harnessed throughput is not more tokens overnight — it is a control plane that refuses to re-pay for context and validation the factory already had.