This work investigates how coding harnesses shape the long‑horizon performance of autonomous coding agents. We built a lightweight harness with a fixed execution loop while swapping out planning, action space, and context‑management components. Experiments on SWE‑Bench Verified and Terminal‑Bench 2.1 covered four models, 176 matched settings, five context‑management strategies, four window budgets, and targeted ablations of planning and action space. Findings include:
- Context management becomes more valuable as the window budget tightens, mainly by preventing overflow failures.
- Staging rule‑based elision before LLM summarization yields the best overall efficiency; making elided content recoverable adds unused machinery and does not improve accuracy.
- Planning acts as an accuracy scaffold for weaker models but serves as a cost saver for stronger ones, with little impact on accuracy.
- Predefined tools help models with limited bash proficiency, while bash‑capable models perform well with a bash‑only interface and achieve substantially lower cost on command‑line‑centric tasks.
Trajectory‑level analysis shows that context management extends execution trajectories without altering agent behavior, planning shifts where trajectories stop, and the action space changes the granularity of code generation. These insights guide model‑ and budget‑aware harness design and provide a modular framework for future component evaluation.
Review