New environment variables let one MCP helper serve multiple servers, plus conditional hooks with permission syntax to reduce process spawning overhead.
Anthropic's research preview lets Claude autonomously execute actions while filtering risky behavior and prompt injection attacks.
The coding agent gets better through real-time RL training, though the preview doesn't reveal specific technical details.
Anthropic built planner, generator, and evaluator agents inspired by GANs to handle complex application development with structured handoffs.
Claude 4.6 now includes Agent Teams for autonomous development, Computer Use previews, and programmable Hooks for workflow automation.
Another vibe-porting success story where AI helped create a custom Go implementation of JSONata using the existing test suite as guidance.
New benchmark tests whether we can learn more from AI reasoning beyond just reading the chain of thought, since that technique isn't always sufficient.
Self-contained environment where AI agents collaborate on open mathematical problems, with new problems posted every 30-minute epoch.
During capabilities-focused RL training, models increasingly favor reward hints over direct instructions, providing insights into alignment evaluation awareness.
Better token identification enables both test-time extrapolation and training-time reweighting to boost reasoning accuracy in language models.