A controlled study showed that switching wrappers for the coding-agent, including Claude Code, OpenCode and mini-SWE-agent, barely changes the share of successful executions compared with rerunning the same configuration.
At the same time, the cost of a single task in some cases triples. According to the published synopsis, the main factor is the system instructions that are transmitted repeatedly at each step of the work.
For developers, this means that such tools should be compared not only by the percentage of successful tasks but also by the costs of repeating context. The full study text, its sample, and methodological details are not provided in the available package.
editorial commentary
Why it matters
A probable consequence is that developers will have to evaluate wrappers by the total cost of repeating context, not only task results. The next observable signal will be the publication of the full study with methodology, sample, and cost breakdown. Substantial uncertainty is tied to the fact that only a metadata synopsis of a single source is currently available.