Study Finds Coding Agents Waste Up to a Quarter of Task Cost on Repeated Work

Coding agents like Claude Code are usually judged on whether they finish a task, not on how much they spend getting there. A paper posted to arXiv and revised on September 28, 2026, turns that around, running the first study focused purely on wasteful spending inside coding agent sessions. The authors examined 1,200 trajectories from Claude Code and Mini-SWE-Agent across four configurations on SWE-bench Verified, then scaled their mitigation tests to 10k trajectories on held-out tasks.

Three habits stood out. Subsumed retrieval happens when an agent fetches information it has already pulled in an earlier step. Similar script generation happens when it writes near-identical helper scripts rather than reusing one. Test re-execution happens when it reruns a full test suite after a change that could not have affected the result. None of these behaviors make the final answer more correct, they just add tokens, and tokens are what the bill is built on.

The paper then tests fixes. Structure-aware retrieval, which gives the agent a map of the codebase instead of letting it search freely, sometimes backfired: it altered how the agent delegated subtasks and pushed cost higher in some runs rather than lower. Agent-synthesized skills, where the model writes its own reusable guidance after a task, produced notes so tied to the specific trace they came from that they rarely helped on a new task. Developer-designed skills, written by a person who understands the general pattern rather than one instance of it, performed best of the three.

For a team running Claude Code or a similar agent at scale, the practical read is narrower than “use better prompts.” It is that the fix with the largest measured payoff required a human to write the shortcut down, not the agent. That is a workflow decision, not a model swap: it means budgeting engineer time to review and author agent skills rather than assuming the agent will learn efficient habits on its own from experience.

The numbers

The study drew on 1,200 trajectories in its main analysis and 10k trajectories in the mitigation tests on held-out SWE-bench Verified and Pro tasks. The three cost-inefficient behaviors combined accounted for up to 22.75% of a task’s total cost. Structure-aware retrieval, meant to cut wasted search, instead raised cost by as much as 28.14% in some configurations by changing how work got delegated. Developer-designed skills cut cost by up to 41.73%, a bigger gain than the agent’s own self-written guidance managed on the same benchmark.

The authors describe the paper as under review for submission, and it covers only two coding agents on one benchmark family, so the exact percentages may shift as other agents and tasks are tested. The direction of the finding, that human-written efficiency guidance beats agent-written guidance, is the part most likely to hold up outside this one study.