Platforms, pipelines and the tooling layer. How engineering teams actually ship software.
A newly open-sourced benchmark from Sierra reveals that leading AI coding agents struggle significantly when tasked with building customer service systems autonomously, with none achieving better than 25% success rates on realistic business requirements.
Sierra's Hyper-τ-bench tests whether AI agents can autonomously build other agents. The best performer, Claude Opus 5 in Claude Code, achieved just 23.9% on the benchmark, revealing systematic weaknesses in information gathering and design exploration.
OpenAI has empowered AI models to reject code submissions that contain security vulnerabilities, fundamentally changing how the company approaches code review, maintenance, and refactoring.
Cloudflare has rebuilt the module registry in workerd to improve standards compliance and Node.js compatibility, enabling developers to deploy larger applications and use more native runtime APIs by default.
Developers can now batch up to 25 code quality findings and have Copilot remediate them in a single operation, with the AI validating changes before opening a pull request.
The Institute of Foundation Models released six open-source AI models ranging from 0.9B to 375B parameters with training code and data, yet some flagship models' transparency remains incomplete at launch.
AI coding agents are overwhelming engineering teams with pull requests, prompting Harness to rebuild its repository from the ground up to handle continuous agent traffic and introducing new code review tooling to manage the deluge.
The latest update to GitHub Copilot for JetBrains introduces enterprise-managed sandbox policies, cross-file navigation for code suggestions, and the ability to connect terminal CLI sessions directly to the IDE.
