Agentic AI
Overview
Agentic AI studies systems that turn goals into sequences of decisions and actions, often combining large language models with tools, memory, feedback, and collaboration with people or other agents. Common approaches include plan–act–observe loops, retrieval and persistent memory, reusable skills, multi-agent orchestration, and runtime harnesses that add verification or policy constraints. Evaluation spans coding and web tasks, proactive assistants, enterprise workflows, and physical or network control, using benchmarks such as GAIA and SWE-bench as well as tests of long-horizon completion, cost, timing, and recovery. Reliable open-world performance remains difficult: errors compound across steps, proactive behavior can overreach, and security, auditability, interoperability, and human authorization must keep pace with autonomy.
Latest papers
215 papers22 papers in the last four weeks, up 5% on the four weeks before. 0.1% of all new papers.