Capable AI Coding Agents

Project information

Title: Capable AI Coding Agents

Author: Robin Rentmeister

Abstract: The use of Large Language Models (LLMs) in software engineering is shifting from single-shot code completion to autonomous, multi-step workflows that operate over large repositories and interact with external tools and APIs. While this transition promises productivity gains, it also introduces concerns for correct and secure results. In contrast to isolated snippet generation, agentic pipelines can amplify small mistakes into cascading failures, exhibit non-deterministic behavior, and produce plausible code that silently violates security or domain constraints.

This survey reviews failure modes in agentic coding, consolidates techniques for improving correctness and reproducibility, and highlights remaining challenges. We organize assurance mechanisms into three complementary layers: (i) operational infrastructure for deterministic execution, provenance, and tracking; (ii) cognitive architectures that couples generation with executable checks and iterative repair; and (iii) interaction paradigms for human and multi-agent interaction, policy enforcement, and governance. Drawing on evidence from execution-based benchmarks and safety evaluations, we summarize which mechanisms yield consistent improvements and where concerns persist. We conclude by outlining open challenges toward scalable, long-horizon robustness and verifiable reliability for AI coding agents.

Agentic AI

All
Team