Agent Harness for Large Language Model Agents: A Survey
An academic survey that formalizes the harness surrounding an LLM — execution loop, tool registry, context manager, state store, lifecycle hooks and evaluation interface — as a primary driver of agent reliability, reviewing over 110 papers across 23 systems.
This repository accompanies a research survey arguing that an agent's execution harness, not just the underlying model, is a major determinant of its reliability at scale. It defines the harness as a six-component tuple (execution loop, tool registry, context manager, state store, lifecycle hooks, evaluation interface) and surveys more than 110 papers, blog posts and reports across 23 agent systems.
The repository organizes this material into a paper list and taxonomy covering historical lineages, security and sandboxing, evaluation and benchmarking, protocol standardization, runtime context management, tool use, memory architecture, planning, multi-agent coordination, compute economics and open technical challenges, alongside a citation entry for the paper itself.