Task-CoEvolve cuts harness eval costs 80%

This is a smart reframing of agent optimization: evaluation becomes an adaptive information-gathering problem instead of a brute-force tax.

• –Variance-weighted sampling focuses compute near the agent’s capability frontier
• –The method reportedly reaches comparable performance with roughly 20% of validation tasks per iteration
• –Terminal-Bench results show 67–80% lower search costs, though runtime savings are smaller because selected tasks can be longer
• –The approach could make iterative prompt, memory, retrieval, and tool-use optimization practical for more teams
• –Reproduction remains limited because the repository currently documents the method while code is still listed as coming soon