Agent systems covers the move from single model prompts to AI that plans, uses tools, and acts across multiple steps. Articles here examine agent architectures, memory, tool use, and the evaluation problems that make autonomous behavior hard to trust. We focus on how these systems are actually built and where current designs still fall short.
Researchers from UNC-Chapel Hill, Carnegie Mellon, UC Santa Cruz, and UC Berkeley built a continual meta-learning framework that gives deployed language model agents two complementary superpowers: instant skill injection from failures, and opportunistic weight-level fine-tuning while you sleep — pushing…
MetaClaw: The LLM Agent That Meta-Learns and Evolves in the Wild Read More »
GraphRAG shreds procedural knowledge into thousands of disconnected fragments. A new framework from Nanyang Technological University puts it back together — and the results in Minecraft are hard to argue with.
Researchers at the University of Calabria built a system that fuses opinion dynamics, consensus theory, and a borrowed trick from solar panel engineering to let networked agents decide—on their own, continuously—which sensors to believe and which to quietly ignore.
A novel test-time framework that intercepts and iteratively rectifies erroneous agent outputs using retrieval-augmented adversarial indicators, achieving 6.3% accuracy improvement on mathematical reasoning benchmarks without model retraining.
AgentDropoutV2: Test-Time Rectify-or-Reject Pruning for Multi-Agent Systems Read More »
A hierarchical framework separates “knowing what” from “knowing how” — enabling co-evolution of declarative planning knowledge and procedural execution skills for mobile device control, achieving 76.1% success rate on AndroidWorld using only raw screenshots.
K2-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control Read More »