Why do coding agents fail on real scientific and software tasks?
The biggest gap is that agents can't handle underspecified workflows—tasks where the instructions are incomplete or the path isn't spelled out. In a 2026 benchmark called AutoMat, researchers tested AI agents on reproducing claims from computational materials science papers. The best-performing agent succeeded only 54.1% of the time [2]. When agents had to reconstruct the workflow from the paper text alone—without step-by-step instructions—their failure rate shot up, mainly due to incomplete procedures, methodological deviations, and execution fragility [2]. This means current agents are brittle: they work well on clean, well-defined coding benchmarks but fall apart when given the kind of ambiguous, real-world tasks that human developers handle daily.
A separate 2025 study of 15,451 AI agent refactorings across 12,256 pull requests found a related pattern: agents overwhelmingly performed low-level, consistency-oriented edits like changing variable types (11.8% of refactorings) or renaming parameters (10.4%), rather than the high-level design changes humans prioritize [1]. The agents' refactorings did yield small but statistically significant improvements—median class size dropped by 15 lines—but the narrow scope suggests agents lack the strategic understanding to tackle complex architectural changes [1]. Together, these studies show that agents are effective at local, well-defined tweaks but hit a wall when tasks require reconstructing context or making design-level decisions.
How does poor compiler feedback hold agents back?
A critical but often overlooked gap is the interface between the agent and the compiler. A 2026 study found that legacy compiler interfaces produce optimization remarks designed for human intuition, not machine logic—they are unstructured, lossy, and often ambiguous [4]. When agents received precise, structured feedback, their success rate tripled (3.3x improvement) compared to ambiguous remarks. Worse, ambiguous feedback actively caused harm: it triggered semantic-breaking hallucinations where the agent made changes that broke the program's meaning [4]. The researchers concluded that the bottleneck is the interface, not the agent itself—small models performed well when given clear, structured compiler feedback [4]. This suggests that improving how compilers communicate with AI agents could unlock significant gains without needing bigger models.
This finding aligns with the broader pattern from the refactoring study [1]: agents are good at executing well-defined local changes but struggle when the feedback they receive is vague or incomplete. The compiler study [4] shows that even a small model can succeed if the feedback is precise, implying that many current failures are due to poor communication between tools, not lack of intelligence in the agent.
What's missing from current testing of coding agents?
Current benchmarks don't test agents on the full end-to-end workflow of real development. The AutoMat benchmark [2] was specifically designed to fill this gap by requiring agents to recover computational procedures, navigate specialized toolchains, and determine whether results support a scientific claim. The low success rate (54.1%) shows that existing software engineering benchmarks overestimate agent capability because they don't test these messy, domain-specific steps [2]. Similarly, the refactoring study [1] found that agents focus on low-level changes partly because benchmarks don't reward high-level design improvements—there's no standard way to measure whether an agent improved architectural quality.
A 2021 study on Cody, a system for semi-automating qualitative research coding, highlights another gap: users want explanations for AI suggestions but rarely use them when provided [5]. This suggests that even when agents could explain their reasoning, the interface or timing isn't right. The study also found that AI suggestions improved coding quality (intercoder reliability rose from 0.085 to 0.33) but not speed, indicating that current agents trade efficiency for accuracy [5]. Together, these studies point to a need for benchmarks that test agents on realistic, underspecified tasks and measure outcomes beyond simple pass/fail.
About These Sources
This answer is built on 5 studies (2 peer-reviewed, 3 preprints) — published from 2021 to 2026, 4 from 2024 or later, collectively cited 88 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 57 papers retrieved from a database of over 500 million.
Sources used in this answer
Agentic Refactoring: An Empirical Study of AI Coding Agents
In the largest empirical study here (15,451 refactorings across 12,256 pull requests), AI agents focused on low-level edits like variable type changes (11.8%) and renaming (10.4%), yielding small but significant structural improvements (median class size reduction of 15 lines).
Can Coding Agents Reproduce Findings in Computational Materials Science?
The AutoMat benchmark tested agents on reproducing claims from computational materials science papers; the best agent succeeded only 54.1% of the time, failing mainly due to incomplete procedures and execution fragility when workflows had to be reconstructed from paper text.
Code with me or for me? how increasing ai automation transforms developer workflows
This paper identifies three goals for building future coding agents, noting limitations in user experience and existing tools, but provides no quantitative results on agent performance.
AI Coding Agents Need Better Compiler Remarks
Precise compiler feedback improved agent success rate 3.3x compared to ambiguous remarks, which actively triggered semantic-breaking hallucinations; the bottleneck is the compiler interface, not the agent's capability.
Cody: An AI-Based System to Semi-Automate Coding for Qualitative Research
In a study of Cody, an AI system for qualitative research coding, AI suggestions improved intercoder reliability from 0.085 to 0.33 but did not speed up coding; users desired explanations but rarely used them.