An Empirical Study of Harness Design for Coding Agents

Which parts of a coding agent's harness actually help depends on the model and the memory budget.

78.7%

of tasks

Unmanaged tasks lost to overflow

At 32k tokens, most unmanaged runs crash before finishing.

How it works

Harnesses tested whole
Agent scaffolds get compared as complete systems, so nobody knows whether planning, tools, or memory handling caused a gain.
Swap one part
A fixed ReAct loop varies planning, bash-only vs tool set, and five context strategies across 4 models and 176 settings.
Effects are conditional
Context management matters under tight windows. Planning rescues weak models but trims cost for strong ones. Strong models prefer bash.

From the paper