Agentic AI Coding: Verified Engineering Over Vibe Coding
A recent paper on arXiv (2605.20456) posits that the primary obstacle in coding for agentic AI has shifted from prompt engineering to process control engineering. While agentic systems are capable of examining repositories, planning actions, modifying files, executing tests, and submitting pull requests, current findings do not substantiate the idea that autonomous code generation inherently enhances results. Controlled experiments indicate productivity improvements in certain enterprise tasks, but also reveal slowdowns in established open-source projects, mixed meta-analytic results, and ongoing challenges with repository setup, dependency management, permission gating, and hardware verification. The paper consolidates insights from various studies on agentic software engineering, GitHub adoption, repository-level configurations, productivity assessments, issue resolution, and hardware/RTL verification.
Key facts
- Paper arXiv:2605.20456 argues engineering process control is the central problem in agentic AI coding.
- Agentic systems can inspect repositories, plan steps, edit files, run tests, and submit pull requests.
- Current evidence does not support that autonomous code generation automatically improves outcomes.
- Productivity gains seen in some enterprise tasks.
- Slowdowns observed in mature open-source work.
- Moderate but heterogeneous meta-analytic effects reported.
- Persistent failures in repository setup, dependency handling, permission gating, and hardware verification.
- Synthesis covers agentic software engineering, GitHub-scale adoption, repository-level agent configuration, productivity trials, issue-resolution benchmarks, and hardware/RTL verification.
Entities
Institutions
- arXiv