Ollama 0.24 ships built-in support for the Codex App, giving developers a local desktop environment for parallel coding threads, built-in browser annotation, and inline code review. A reworked MLX sampler also improves generation quality on Apple Silicon.

A new study finds that LLM coding agents degrade sharply as structural constraints accumulate in backend code generation tasks. Agents that perform well under loose specs often collapse when architecture, ORM, and framework conventions all apply at once.

AI agents like Claude are confidently agreeable, which makes them dangerous in architectural roles. A consultant's field report explains what goes wrong and who ends up holding the bag.

OpenCompass 0.5.2 ships fourteen new benchmarks spanning science, biology, chemistry, and instruction-following, plus new model and API support. Here is what product engineers need to know to put it to work today.

Dify v1.14.2 tightens tenant-scoped security, restores workflow tracing after human-in-the-loop resume, and fixes a cluster of knowledge-base and RAG reliability bugs. Here is what changed and what it means for teams running Dify in production.

CowAgent 2.0.9 ships MCP protocol support, a unified Models page in the Web console, and persistent browser sessions. Builders can now plug any MCP-compatible service directly into the agent as a tool.

Hugging Face Transformers v5.9.0 adds Cohere2Moe, Parakeet TDT, and HRM-Text to the library while breaking SAM3 and EdgeTAM input contracts. Audio support also gets a notable expansion.