Meta researchers have developed a structured prompting technique — called semi-formal reasoning — that enables LLMs to verify code patches without executing them, reaching up to 93% accuracy on real-world agent-generated patches. The technique fills a practical gap between free-form chain-of-thought reasoning (flexible but prone to hallucinations) and rigid formal verification (precise

When a coding agent fails, the instinct is to examine the model or the prompt. A new paper argues that this instinct misses half the problem: real-world agent failures split roughly evenly between tool-use accuracy (how well the agent invokes a tool) and intrinsic tool accuracy (whether the tool itself

Training effective software engineering agents requires large volumes of high-quality task trajectories, and generating them at scale is expensive. A new paper extends the "Less-Is-More" (LIMO) hypothesis — which showed in mathematical reasoning that a small number of excellent examples dramatically outperforms large volumes of mediocre ones — to the

The largest real-world study of AI-assisted coding in IDE-native settings to date has landed, and its findings challenge some of the most common assumptions about how developers actually use these tools. Researchers analyzed 74,998 developer messages across 11,579 chat sessions — all drawn from real sessions committed to public

A new large-scale empirical study has done something the agentic coding field has been waiting for: it went outside the benchmarks entirely and examined 110,000 real open-source pull requests from five production coding agents — OpenAI Codex, Claude Code, GitHub Copilot, Google Jules, and Devin — to measure what actually happens