The Agent Harness: An Agent Is More Than a Loop

This post examines AG2’s beta Agent architecture, describing six pluggable layers (“harness”) including MemoryStream, AssemblyPolicy, Middleware, Tools, Observers, and KnowledgeStore. It argues an agent has moved past a simple LLM/tool-call loop, now managing memory spanning thousands of turns, and closes with a production example combining token monitoring, loop detection, and persistent memory.