FeatureArc: Coding agents fail 55% on multi-stage features
Benchmark reveals steep decline in agent coding performance across feature stages
A newly published research paper titled *FeatureArc: Benchmarking Coding Agents on Evolving Features* has exposed a critical weakness in current coding agents: their inability to maintain code quality as software requirements evolve. Authored by Wang Yunlong, Alyssa Wu, Cuiyun Gao, Jia Feng, Songyuanyi Lu, and Zhichao Hu, the paper was submitted to arXiv on October 27, 2026, and introduces a systematic evaluation framework for testing how agents handle incremental feature development arXiv.
The research evaluated 10 coding agents across 73 manually reviewed tasks spanning feature sequences of 2 to 10 stages and covering four programming languages. The findings paint a stark picture of agent limitations: while agents achieved a pooled success rate of 62.7% on initial feature stages, this rate collapsed to just 28.5% when evaluating the complete feature against final code arXiv. This represents a 55-percentage-point drop in performance—a dramatic failure cascade as complexity accumulates.
Inherited defects create compounding failures
The root cause of this performance cliff lies in how agents handle errors across sequential stages. Inherited defects were present in 72.9% of failed later stages, indicating that agents rarely remediate mistakes introduced in earlier development phases arXiv. Most troublingly, after an initial failure, agents left the earlier error unnoticed in approximately 95% of subsequent stages, effectively blind to technical debt they had created arXiv.
This pattern has immediate implications for real-world deployment. Coding agents marketed as autonomous developers would accumulate defects silently, requiring expensive human review and refactoring at each stage. The benchmark demonstrates that current agents lack the reflective capacity to audit their own prior work—a fundamental capability gap that distinguishes human developers from machine agents.
Implications for agent product roadmaps
The FeatureArc findings challenge vendors positioning coding agents as drop-in replacements for junior developers. Products that promise to own complete feature development from specification to shipping code would, based on this benchmark, deliver increasingly broken implementations as features grow in scope. Teams deploying such agents would need to establish human-in-the-loop review at every stage—negating productivity claims.
The paper's use of manually reviewed tasks across real programming languages strengthens its credibility against claims that agents perform well on synthetic benchmarks but fail on authentic work. With the benchmark now public, vendors have a measurable standard against which to evaluate and improve their agents' cross-stage consistency and error recovery mechanisms.
For the agent developer community, FeatureArc establishes a research baseline for tracking improvements in multi-stage reasoning and code quality maintenance—core requirements for agents to handle enterprise software development workflows.