Beyond the Model: A Survey of Agent Harnesses for Long-Horizon LLM Agents

PhD Qualifying Examination


Title: "Beyond the Model: A Survey of Agent Harnesses for Long-Horizon LLM
Agents"

by

Mr. Zhongwei XIE


Abstract:

As LLM-based agents advance from static, single-turn question answering to 
long-horizon autonomous systems, their performance increasingly depends on 
the surrounding infrastructure, broadly referred to as the Agent Harness. 
Yet, the Agent Harness remains difficult to characterize, spanning 
heterogeneous mechanisms for execution, memory, and interaction. Existing 
surveys typically study these mechanisms in isolation, rather than as an 
integrated agent infrastructure. To bridge this gap, we introduce a layered 
architectural taxonomy that organizes the Agent Harness into four 
interdependent layers: (1) Execution & Orchestration, (2) Context & 
Trajectory Management, (3) Interaction Surface & Execution Environment, and 
(4) Constraints & Guardrails. Using this taxonomy as an analytical lens, we 
examine how harness-level mechanisms shape agent capabilities across existing 
systems. This analysis yields two key findings: the Agent Harness is a widely 
overlooked performance bottleneck, and it co-evolves with the underlying 
models. These findings suggest that future harness architectures should be 
designed for change, with modular boundaries that allow obsolete scaffolding 
to be removed or replaced as model capabilities advance.


Date:                   Tuesday, 6 October 2026

Time:                   1:45pm - 2:30pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Dr. Yangqiu Song (Supervisor)
                        Dr. Binhang Yuan (Chairperson)
                        Dr. May Fung