Code & Chain · Signal Desk

Tests Show Nearly Three-Quarters of AI Coding Agent Tasks Still Need Human Finishing

Original sourcethe-decoder.com

Summary

A new benchmark, Co-Design Bench, tested coding agents with real users' web-app redesign requests, including image, audio and document inputs. Results showed that even the strongest model, GPT-6 Astra, got only about 26.4% of tasks fully correct, with average task coverage around 70%, but nearly th…

Key points

  • Provides testable expectation-management data using real redesign requests, avoiding overestimating fully automated development.
  • The data directly illustrates the gap in agent coding; enterprises should validate on a "model plus environment" basis when adopting.
  • When estimating adoption benefits, teams should factor in human review and cleanup hours and benchmark their own harness rather than relying only on model rankings.

Editorial note

This page is Code & Chain's editorial summary of public sources. It may be prepared with AI assistance and published through an automated workflow. Refer to the original sources; this content is not investment, legal, or tax advice.