Agent ecosystem & evaluation methodology explosion

New benchmarks: ASI-Bench (autonomous science), AxiomProver (math milestone), SWE-Atlas, VideoGAIA, R^3-Bench. DeepSeek Harness open-sourced. Security: DeepSeek Harness prompt injection (25.5% success), mind viruses. Methodological advances: uncertainty propagation, harnessed agentic RL, evolution strategies for fine-tuning. Lab-to-production gap remains 37%.