Open-source maintainers have coined a term for what AI coding agents are doing to their contribution queues: "AI slop." Low-quality AI-generated code, PRs, documentation, and bug reports are flooding development workflows at a pace that review processes weren't built to handle. A new qualitative study puts

Coding agents are often described as risky for production codebases, but the risk is not uniformly distributed. A new comparative study of 7,191 agent-generated pull requests and 1,402 human-authored PRs from Python repositories in the AIDev dataset finds a counter-intuitive split: coding agents introduce fewer breaking changes than

Most research on AI-generated code quality operates under controlled conditions — short-lived experiments, synthetic benchmarks, carefully curated prompts. What happens in production repositories over time is a different question, and until now it's been largely unanswered. A new large-scale empirical study changes that by tracking 304,362 verified AI-authored

The default pattern for coding agents exploring an unfamiliar codebase is expensive and surprisingly shallow. When an agent needs to understand what a function calls, where a class is used, or which modules depend on each other, it falls back on repeated file reads and grep searches — accumulating thousands of

AGENTS.md and CLAUDE.md have become standard practice in AI-assisted development — the go-to mechanism for telling a coding agent what it needs to know about your project. A controlled study across real repository benchmarks now complicates that assumption. Developer-written context files produced a modest performance improvement. LLM-generated context files

A supply-chain team deployed two AI agents simultaneously — one optimizing procurement speed, one minimizing daily capital spend. When demand spiked, the agents entered complete gridlock: the procurement agent generated purchase orders; the finance agent canceled them immediately. The cycle repeated thousands of times, burning compute tokens while delivering zero operational

Long-horizon agents — those capable of working through multi-step tasks over extended sessions — are no longer a research prototype. They are operational in real engineering environments today. But operational does not mean autonomous. The meaningful breakthrough of early 2026 is not that agents can now be left unsupervised; it is that

Vibe coding's most common real-world use case — "build me a website" — has had no rigorous benchmark until now. Vision2Web, accepted at ICML and built by Zehai He et al. at Zhipu AI, fills that gap with 193 tasks, 918 prototype images, and 1,255 test cases

JetBrains just announced Central, an open platform for unified AI agent management — and the timing reflects a specific operational crisis that's only visible once a team gets past the "our first agent works" milestone. The problem isn't agent capability. It's what happens

Most teams running web-augmented coding agents — those that issue live search queries before generating code — have no systematic defense against bad search results quietly degrading output quality. A new paper from Guoqing Wang et al. documents exactly what happens when that defense is absent, and introduces Sherlock, an automated pipeline