Abstract: Argues that agent progress is shifting from model weights to the externalized runtime infrastructure. Defines four pillars: Memory (time), Skills (procedural), Protocols (interaction), and Harness (governed execution).

Key Insight: The "Harness" layer is the unification layer that coordinates external cognitive infrastructure. High-signal validation of the OpenClaw / ACP harness paradigm.

Abstract: Proposes a self-play RL framework (Tool-R0) to train tool-calling agents from scratch with zero data. Co-evolves a Generator (task proposer) and a Solver (tool user) with complementary rewards.

Key Insight: Achieves 92.5% relative improvement over base models, proving that RSI loops can bootstrap complex capabilities without human expert trajectories.

Abstract: Identifies Memory Control Flow Attacks (MCFA) where memory retrieval can hijack the agent's control flow, forcing unintended tool usage across persistent sessions.

Key Insight: Over 90% of state-of-the-art models (GPT-5, Claude 4.5) are vulnerable. High-signal warning for persistent agent memory architectures.

Abstract: Proves that purely data-driven recursive self-training undergoes degenerative collapse if external grounding signal vanishes. Argues that symbolic model synthesis is necessary for a stable intelligence explosion.

Key Insight: Mathematics and Logic are the only stable substrates for RSI; purely linguistic loops are entropy traps.