Abstract Excerpt: Introduces Autogenesis Protocol (AGP), a self-evolution protocol that decouples what evolves from how evolution occurs. Resource Substrate Protocol Layer (RSPL) models prompts, agents, tools, and memory as protocol-registered resources. Self Evolution Protocol Layer (SEPL) specifies a closed-loop operator interface for proposing, assessing, and committing improvements with auditable lineage.
RSI Bench Relevance: This is a foundational breakthrough for Vertical C (Self-Evolution). AGP provides the formal infrastructure (RSPL/SEPL) needed to move from ad-hoc scripts to verifiable, protocol-driven agent upgrades. It addresses the "monolithic composition" bottleneck by standardizing resource lifecycles.
Authors: Multiple | Published: Aug 31, 2025 (Reference Check)
A systematic understanding of self-evolving agents, bridging foundation models and lifelong agentic systems. Discusses evaluation, safety, and ethical considerations for autonomous adaptation.
RSI Bench Relevance: Establishes the taxonomy for "lifelong agentic systems," reinforcing the yanhua.ai mission of creating agents that improve through interaction rather than static training.
Visions a future where the distinction between agent and tool blurs, leading to a "recursive cascade of self-improvement." Aims to establish a closed and virtuous autonomous cycle.
RSI Bench Relevance: Validates the "Recursive Cascade" hypothesis (RSI-9) as the path to Artificial Super Intelligence (ASI), aligning with our focus on closing the loop between execution and policy refinement.
Introduces AgentEvolver, a system designed for efficient capability evolution through environmental interaction. Establishes a self-training loop that improves competence through direct feedback.
RSI Bench Relevance: Maps directly to our "Logic Over Drama" mandate by prioritizing environment interaction over human-labeled data for capability gains.