I just have custom instructions to tell the model to edit files using its tools rather than a gigantic bash pipelines. That's 80% of it.

▲

siraben a day ago | parent | next [-]

I prefer not the fight the RL as the models get better, unless I've measured it and haven't observed any degradation. There may be other costs/no. of turns considerations too.

▲

machomaster a day ago | parent | prev [-]

Would that be useful in other harnesses/models like Codex? Wouldn't that lead to increased token usage?

	▲	alex_duf 6 hours ago \| parent [-]
		I assume it works with any harness. Frontier models have been pushed to reduce their token usage so much that they got to be creative and use bash CLI to edit files, which achieves the goal of reducing token usage but makes reading the traces less practical. It's a trade-off. Personally I like occasionally re-reading the edits and I prefer use more tokens and understand what's happening (or have the illusion to understand).