Artificial Intelligence Code Writers Millions of Lira Empty

Although artificial intelligence-based code writing agents speed up the development process, the research presented reveals that these systems have serious financial inefficiencies. Science people Claude Code and Mini-SWE-Agent have identified patterns that improve their behavior in detail.

Results made within the scope of research:

-1.200 code writing sessions were reviewed — Claude Code and Mini-SWE-Agent were analyzed in four different configurations of SWE-bench Verified platform.

• Three main expensive behavior detected — Unnecessed data screening (subsumed retrieval) significantly increases the costs of similar script repetition and re-run of test processes.
• Benzer script files are regenerated — Agents do unnecessary API calls by creating similar solutions to the code they write in advance.
• Tests are run on unnecessary location — The same test scenarios are repeated multiple times and this significantly increases the cost of tokens.

These findings show that companies are an additional layer of cost when using artificial intelligence codes. Research emphasizes that developing strategies for mitigation of these expensive behavior is an urgent requirement. The next generation of code writing systems must be designed by taking into account financial efficiency.