Co Design Bench Coding Agent 26 Percent

The best coding agent tested fully completes only 26.4% of real user tasks.

A new benchmark called Co-Design Bench drew 273 tasks from 26,314 real coding sessions. The top agent met about 70% of the acceptance criteria on a typical task. It fully finished about one task in four.

That gap is the story. Partial credit looks impressive in a demo. In production, a task that is 70% right still needs a person to find the missing 30% and fix it.

Older benchmarks reward narrow, well-defined problems, and agents score very high on them. Real sessions are messier, with vague requirements and context spread across files. That is the work your team actually does.

My view: the metric to ask vendors for is full-task completion on work that looks like yours, not benchmark scores. The review burden is where the cost hides.

If you are budgeting for coding agents, plan for human review of every output until your own data says otherwise.

What completion rate would an AI coding agent need to reach before you trusted it without review?