Do Coding Agents Really Fail at Business Logic? What the Evidence Shows

The claim that coding agents struggle with bug fixing due to a lack of understanding of business logic is supported by some studies but appears narrower in scope than initially thought. Research involving Claude Code and multiple agents shows that bug fixes are merged less frequently than other types of work, such as documentation and performance tasks. However, about 64% of rejected pull requests (PRs) were due to reviewer abandonment, while only about 2% were caused by issues with the agent itself.

Another study focused on bug-fix PRs alone found that 46.41% of fixes from Copilot, Devin, Cursor and Claude received rejections. The reasons were grouped into four categories, with "wrong implementation" being the most common. After a fix was merged, code smells were the most common issue, although real bugs were rarer but often severe. This suggests that even when agents do fix a bug, there may still be quality issues that need to be addressed.

However, the studies do not measure whether a fix violates a business rule, which means the actual impact of the problem on business logic is indirect. Developers have reported similar patterns, but these are anecdotal and not backed by measurements. Still, some developers find agents to be helpful in debugging complex issues across multiple files, while others note that agents can struggle with understanding the broader context of a change.

Existing tools focus more on navigation, impact analysis and workflow rather than specifically checking if patches respect business rules. While some tools like Repo Mind and GroundGraph attempt to provide a repository understanding and link requirements to code, they are still in early stages and have limited adoption. There is no existing tool that captures business rules that exist only in people's heads, which is crucial for ensuring that patches meet business requirements.