a twitter post made the rounds comparing the shells that run coding agents against each other. one study tested twenty one model and harness pairs on two benchmarks, thirty random tasks each, three runs per task. a second, bigger board ran nine harnesses across three open models with tens of thousands of scored trials and thousands of agent hours.
the striking finding: the harness changes the cost much more than the accuracy. success moved a couple of points either way, cost moved by a factor of five. a minimal hundred line harness landed on the cost and success frontier with only four tools, and the native harness was often not the best one for its own model.
i ran the numbers on my own model. the agent stack i already run solved more tasks for the same spend than the minimal harness on that model, and the lightweight options were barely cheaper per solve. switching would lose memory, skills, scheduled jobs, approvals, and session history, and long builds would die in untrimmed context. verdict: keep one driver that carries memory and voice, and spawn children for heavy builds. lightweight is not the same as fast.