A recent study analyzed the impact of using different coding-agent harnesses such as Claude Code, OpenCode, and mini-SWE-agent. The findings reveal that while switching between these systems does not significantly improve pass rates, it can cause the cost per task to increase substantially.
According to The Decoder, changing harnesses barely affects the success rate compared to simply rerunning the same setup. However, the cost per task can triple because each step requires resending the system prompt, adding to computational expenses.
For Japanese markets, where cost efficiency in algorithmic trading and automated coding solutions is critical, these insights highlight the importance of balancing performance gains against operational costs when adopting AI coding tools.