Google Cloud AI Research open-sourced RRSI, a framework that allows language model agents to rewrite their own prompts, tools, and memory while keeping model weights frozen, reducing overfitting through a leakage critic, noise floor, cost rule, and pruning mechanisms (MarkTechPost). Testing with Claude Opus 4.8 on Terminal-Bench 2.1 showed performance gains that carried over to new tasks. The release addresses a core challenge in agent development: improving reasoning and tool use without expensive model retraining.
H Company released Holo4, a family of open-weight generalist computer-use models available in two sizes - Holo4 27B dense and Holo4 35B-A3B mixture-of-experts with 3B active parameters - supporting screen interaction, code generation, and MCP or API tool calling across desktop, web, Android, and APIs with 256K context windows (MarkTechPost). In related agent infrastructure work, Google Research’s RRSI framework emphasizes source attribution for MCP agents to improve verification accuracy beyond factual correctness (Hugging Face Blog).