Berkeley Paper: Agent Harness Choice Can Cut Costs 2.6x

A UC Berkeley paper shows that agent harnesses are far from neutral: with the same model and task, optimizing the harness cut costs up to 2.6x, and different frameworks differed twofold in token usage on the same repo bug-fixing task.

2026-11-11 ~ 2026-11-11 · 2 related posts

• Same model, half the tokens: benchmark shows the agent harness is not neutral plumbing — uriwa · 2026-11-11
• Berkeley Paper: Harness Tweaks Cut Agent Costs Up to 2.6x Without Changing Models — dair_ai · 2026-11-11