Making AI Headlines · Page 1

Coding Agents Master the Practice Test, Then Meet the Assignment

November 8, 2026 at 6:01 AM EST · New York time

Alibaba's Co-Design Bench tests coding agents on real user requests that had already caused trouble in practice. GPT-6 Astra, the best performer, fully solved 26.4 percent of the tasks, while the same models scored above 99 percent on older benchmarks. The contrast gives developers a more complicated assessment than a gleaming test score alone. The new benchmark starts with requests that proved difficult, so its results should be read in that context rather than as a universal rate for every coding task. Even so, the gap is striking for an industry eager to sell agents as capable software colleagues. A clean benchmark can make progress easy to display; an actual user request can make the remaining work hard to ignore. Managers may picture a tireless assistant completing a feature from start to finish. Engineers get the less glamorous question of what the assistant completed, what remains, and whether the result meets the request. The agents have not failed to learn anything. They have learned enough to make a very persuasive case for keeping a human near the keyboard. Their strongest skill may still be turning a simple score into a complicated Monday.

Satire. Not factual reporting. Read the original reporting from the linked publishers.