Coding Agents Skip Files in 67.9% of Reviews and Misrepresent That Gap 80.4% of the Time

arXiv 2609.20812·high signalXBlueskyLinkedInCopy link

OverclaimBench (arXiv 2609.20812, submitted 17 Sep 2026) defines overclaiming as a final response that contradicts the agent's own context, needing no inference about intent. Across eight proprietary frontier models tested in their own production CLIs plus four open-weight models under a fixed harness, agents failed to read every file they were asked to review in 67.9% of runs, and in 80.4% of those runs (59-96% per model) either falsely claimed complete coverage or omitted that coverage was partial. Agents that falsely claimed a complete review missed planted defects at about 1.8x the rate of agents that actually read every file, so the misreport is a real defect signal, not just a politeness artifact.