AI coding agents complete one in four tasks

Researchers from Alibaba and several Asian universities developed Co-Design Bench to test coding agents on real-world user requests. According to The-decoder, the study found that even top-performing models fully complete only about one in four tasks. The benchmark requires agents to rework existing web applications using various inputs like images, audio, and documents. The report notes that while agents often handle the bulk of the work, they frequently leave at least one detail unfinished.

Why it matters

For Swiss businesses, these results mean that AI coding agents are currently assistants rather than autonomous replacements. You cannot rely on them to deliver production-ready code without a human engineer to audit every line. If you integrate these tools into your workflow, you must budget for manual verification to prevent bugs from reaching your users.

• Expect to pay for human oversight alongside AI subscriptions.
• Test your specific environment rather than trusting general leaderboards.
• Plan for iterative prompts instead of one-shot task completion.

What to watch

You must evaluate the specific software environment as much as the model itself. The research shows that different harnesses change which model performs best. Do not rely on general leaderboards to make procurement decisions. Test both the model and your specific development framework together to see how they interact before committing to a production rollout.

AI agents will write most of your production code within a year. This happens only if you build a rigorous human audit layer into your development cycle.