Harness Tweaks Change Agent Cost 2.6x, Accuracy Barely Moves

Your agent bill keeps climbing, so your first instinct is to downgrade the model. A new paper from UC Berkeley researchers says that is probably the wrong lever.

On the same model and the same task, changing only the harness — the prompts and the tools around the model — barely moves accuracy but shifts cost by up to 2.6x.

What Happened

The researchers measured how much of an agent's bill comes from the scaffolding around the model rather than the model itself. Prompts and tools turned out to be close to free in accuracy terms and very expensive in dollar terms: same model, same task, and the cost still swung by as much as 2.6x depending on how the harness was set up.

To test fixes without rewriting each setup by hand, they built Agency, a layer that sits between existing harnesses like Claude Code and Codex and the model underneath. Because the layer is shared, an optimization written once carries across harnesses instead of being rebuilt for each one.

They then ran five proof-of-concept techniques through it. Those cut cost by up to 35%, with what they describe as marginal accuracy loss.

Why It Matters

If your agent bill keeps growing, look at repeated context reads and extra tool calls before you switch models. The model is often not what changed — the harness is, and that is where the cheaper wins are sitting.

The 35% figure comes from proof-of-concept techniques, not a shipping product, so treat it as a direction to check rather than a guarantee.