RELAI-VCL compounds optimization gains on Terminal-Bench 2.0

A study evaluates whether agent-optimization gains compound over time by testing three methods—GEPA, Meta Harness, and RELAI-VCL—on the hard tasks of Terminal-Bench 2.0. While all methods improve upon a static baseline, only RELAI-VCL successfully transfers to unseen tasks and continues improving after subsequent optimization rounds.

• GEPA's optimized agent performed below the unoptimized baseline when new tasks were introduced.
• Meta Harness transferred well but failed to improve further with a second optimization budget.
• RELAI-VCL achieved the highest lifelong average pass rate at 76.4%, compared to 66.0% for GEPA, 64.6% for Meta Harness, and 58.7% for the baseline.

The authors conclude that optimization gains compound only when regression control is built into the loop, providing an inductive bias against shortcut solutions.