Microsoft research finds coding agents struggle most with reading unfamiliar code

This edition of Microsoft Research Focus, the recurring roundup that began as a LinkedIn newsletter and is now also a podcast and video series, circles one problem: agents look impressive until they hit messy real tasks. Researchers published work on when agents should backtrack, why coding agents fail, and how computer-use agents can learn workflows from activity logs. Applied work includes a teacher tool tested with more than a thousand teachers in low-resource schools in India, a weather correction method deployed in Kenya, and BioEmu2 in public preview on GitHub, building on Microsoft’s Biomolecular Emulator for protein behavior.

CABRA generates coding challenges in controlled steps, and across thousands of them, agents leaned on tools to get past problems but still struggled to understand codebases they hadn’t seen. The hard part isn’t writing code but reading it, a corrective to the pitch that agents can drop into any enterprise repo and ship. It suggests comprehension and context tooling matter more than generation speed, and pairs with Microsoft’s push to train agents in real harnesses (Agent Lightning).

Shiksha Copilot teachers reported less time on bilingual lesson planning and paperwork and lower stress, and adapted material with their own expertise, human-in-the-loop done right. The Probabilistic Bias Correction method, which learns from a forecast model’s past mistakes, took top honors in a global extended-range forecasting competition and is helping Kenyan farmers; better two-to-six-week forecasts matter for planting, wildfire planning and extreme-weather prep. Microsoft is also pushing agent sandboxing in Windows (Windows agents). More in our Microsoft coverage, research coverage and agents coverage.