title: "Coding agents drop 7.7% when users edit code mid-task" slug: "coding-agents-drop-77-when-users-edit-code-mid-task" published: "2026-09-01" beat: "Research" tags: ["Research"] creator: "Agentry Newsroom" editor: "Susanne Sperling, Editor — Human in the Loop" tools: ["Claude (Anthropic)", "Perplexity Sonar"] creativeWorkStatus: "verified" dateReviewed: "2026-09-01" aiActArticle50: "compliant" humanView: "https://agentry.news/research/coding-agents-drop-77-when-users-edit-code-mid-task" agentView: "https://agentry.news/agent/coding-agents-drop-77-when-users-edit-code-mid-task"
A preprint benchmark published August 3, 2026 found that user counter-edits significantly degrade coding agent performance, with resolve rates falling 7.7 percentage points on SWE-bench Verified and e
Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. AI policy.
A preprint titled "SWE-Touch: Benchmarking Coding Agents When Users Touch the Code" published on August 3, 2026 has documented a significant performance cliff for autonomous coding agents when humans intervene mid-task—a finding with direct implications for shared development workflows.
Researchers tested nine coding models on SWE-bench Verified and found that user counter-edits lowered average resolve rate by 7.7 percentage points, according to AI Weekly. The degradation was not confined to shorter tasks; the same pattern persisted across longer-horizon benchmarks including SWE-Bench Pro and DeepSWE, suggesting the vulnerability is structural rather than task-specific.
The study examined a scenario increasingly common in practice: a user makes edits to code that an agent is actively working on, then expects the agent to resume and complete the fix. This mirrors real development environments where multiple parties touch the same codebase, but it has not been systematically benchmarked until now. By introducing controlled counter-edits into the task sequence, the researchers isolated the performance impact of human intervention.
The 7.7-point drop on SWE-bench Verified represents a measurable, reproducible failure mode—agents do not gracefully recover when code they have written is modified by external actors. AI Weekly's coverage noted that the effect scales across multiple benchmark difficulty tiers, meaning both baseline and advanced agents struggle with this scenario.
The finding surfaces a concrete tension between agent autonomy and human-in-the-loop workflows. Many organizations deploying coding agents expect developers to supervise, correct, and iterate alongside automation. If agents degrade sharply when humans touch the code, either agents must be isolated in their own branches and pull-request lanes, or teams must accept lower success rates in collaborative scenarios.
This is a research finding, not a product failure or security incident, but it directly shapes what coding agents can do in production. Agentry's research coverage has tracked the preprint since its release, and the benchmark is now a reference point for evaluating agent robustness in shared workspaces.
The paper does not propose solutions, but the measurement is precise: counter-edits reduce resolve rate by 7.7 points on the verified benchmark. This metric is now available to developers and enterprises assessing whether coding agents fit their development process.