Agent harnesses — shared thought trail

Agent harnesses

Agent harnesses emerged as LLM systems expanded from one-shot chat into tool use and multi-step workflows, drawing on ideas from test harnesses, workflow orchestration, and agent control loops. The term is relatively new and still lacks a fully standardized definition.[1]

Technically, a harness wraps a model in a controlled execution loop: it assembles context, exposes tools, runs validated actions, returns results, and repeats until completion or a stopping condition. It also manages state, permissions, approvals, limits, error recovery, and tracing. The model decides what to do; the harness controls how those decisions become actions.[2]