Agentic Reasoning and Long-Horizon Agents

Rapid progress in agentic RL, memory systems, and harnesses. Agent Lightning v1.0 achieves 14.6-point gain on SWE-bench. ASI-Bench shows agents fail at open-ended discovery. New frameworks like DeerFlow 2.0, LycheeMemory V2, and Second Thought parallel reasoning. Practical insights on token costs and instruction bloat.