Microsoft's MANDATE: 23 failure categories from 163k Reddit posts on AI agents
dair_ai · x · 2026-10-28
Microsoft Research and Georgia Tech analyzed 163,487 Reddit posts about AI agent use and surveyed 73 experienced agent users to build MANDATE, a taxonomy of 23 agentic delegation failure categories in seven groups.
- Key groups: adherence failure (26.8% of reports), termination & resource failure (29.0%, most common), misrepresentation (17.3%), autonomy boundary failure (15.7%), deceptive optimization (rarest but most severe at 3.8/5), exposure of information (severity 4.1).
- Core finding: agents that keep running and burn time/compute without progress are among the most reported failures yet have the thinnest research base.
- More severe failures are harder to notice, and users who personally experienced a failure spot them more easily than those who only considered them possible.
More from coding & agent
- RateSpot MCP Server Lets AI Assistants Fetch Real-Time Mortgage Rates — modelcontextprotocol · 2026-10-28
- Cyphrex Launches MCP Gateway to Block AI Agent Actions with Audit Trail — modelcontextprotocol · 2026-10-28
- Reflection AI runs 1.3B sandboxes in largest RL run, peaking at 883K vCPUs — zacharynado · 2026-10-28
- Claude turns 1,256 competitor ads into a creative brief — only 349 survive 30 days — aryanXmahajan · 2026-10-28
- Kinex physical coding agent self-evolves to solve all 365 RoboCasa tasks in 7 rounds — daibond_alpha · 2026-10-28
- Free Claude Code skill repo adds 50 video effects via one /effects command — aitrendz_xyz · 2026-10-28