Coding AI

A new study finds that code-repair models often rewrite more than necessary, even when their patches pass tests. Preservation prompts and reinforcement learning can improve the fidelity of their edits.

A large experiment finds that coding agents often disagree on third-party services. Repository language, project context and the way vendor information is presented can all change the final choice.

PaperCompiler addresses a central weakness of paper-to-code systems: generated repositories can lose algorithmic details and consistency even when the code appears runnable. Its approach compiles implementation evidence from a paper into traceable, repository-wide specifications before code generation begins.

A new study presents an end-to-end specialization pipeline for competitive programming, combining curated problems, synthetic reasoning traces, post-training, and iterative test-time refinement. The reported system scored 535.4 out of 600 in an IOI 2026 evaluation.

Harness-of-Harness is a framework built on existing coding-agent harnesses rather than a new code model. It organizes autonomous work into repeated planning, coding, testing, and evaluation cycles.

A dispute over one instruction-file format has become an enterprise tooling and governance issue. Shopify’s CEO considered banning Claude Code internally, while Anthropic’s response left developers unconvinced.

LinkedIn built a multi-agent review platform instead of simply placing a general-purpose AI reviewer in front of its repositories. The system combines independent reviewers, repository-specific rules, filtering, and production-grade operations to improve signal quality and measure whether developers actually adopt its findings.

Uncle Bob is experimenting with handing implementation to AI agents and moving human effort toward automated quality controls and architectural review. The approach reduces routine inspection, but it does not replace judgment about system structure, security, or business context.

LEGO-RL addresses a practical gap in training coding agents: native harnesses provide realistic tools and execution feedback, but their behavior can be difficult to align with policy-gradient optimization. The framework improves Qwen3.5-35B-A3B on SWE-bench Verified across three agent harnesses.

Oracle has barred AI-generated code from OpenJDK contributions while still allowing developers to use LLMs privately for debugging and review. The move contrasts with the company’s public claims that AI is reshaping its own software development.

Prime Intellect has open-sourced Prime Agent, a programming agent harness centered on Recursive Language Model and Continual Harness abstractions. Paired with Opus 5, it reportedly reached 95.5% RHAE Best@1 on ARC-AGI 3, slightly above the human expert baseline cited by ARC.

A new study argues that many AI-generated patches are not merge-ready because models often keep code that should be removed.

Uber’s experience shows that once AI coding tools move from experiments to everyday engineering workflows, the hard problem shifts from adoption to cost, quality, and infrastructure sustainability.

FrontisAI introduces OpenMLE, an open full-stack system for studying AI systems that improve the AI development process itself. On top of it, the team post-trains Frontis-MA1, a 35B meta-evolution agent for machine learning engineering.

A new paper studies how reinforcement learning can optimize generated code for execution speed as well as correctness. The key finding is that timing-based rewards are fragile unless testing, reward design, and GRPO training are redesigned together.

Kimi Code now recommends k3-256k, a 256k-context version of its Kimi K3 coding model. It is positioned for routine development work while using less quota than the 1M-context K3 option.

A study of 1.02 million reviewed pull requests across 207 GitHub projects finds that AI agent involvement is linked to faster review decisions, but not consistently better review quality.

Amazon Web Services has released Claude App Gateway for AWS, a self-hosted control plane for Claude Code and Claude Desktop. The gateway centralizes access control, policy enforcement, routing, telemetry, and spending limits for enterprise deployments.

OpenAI’s first branded hardware product is not a phone or an AI companion, but a palm-sized programmable keyboard for Codex users. Built with Work Louder and priced at $230, Codex Micro is designed to control AI coding agents.

Linus Torvalds has made clear that Linux will not become an anti-AI project. The debate now centers on how open source communities should balance useful AI assistance with maintainer burden and code quality.

A new arXiv study of 2,361 popular GitHub repositories finds that agentic coding tools are entering open-source workflows, but intensive use remains concentrated in a small number of projects.