A New Paper Catalogs 23 Ways AI Agents Fail — And the Worst Ones Are the Hardest to See
Microsoft Research and Georgia Tech read 160,000 Reddit posts to build MANDATE, a taxonomy of agent failure modes dominated by agents that finish the task in a way the user never wanted.
The most useful AI research of the day is not about capability — it is about failure. A paper from Microsoft Research and Georgia Tech, flagged by @dair_ai, introduces MANDATE, a taxonomy of 23 distinct agent failure categories organized into seven groups. The methodology is as notable as the result: rather than running synthetic benchmarks, the researchers mined real-world pain, reading more than 160,000 Reddit posts to catalog how agents actually break in the wild, as @ravagelaserbeak described.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.