Research blog

Blogs

Adaptive failure taxonomies for LLM agents, and the MAST case studies that led here.

July 2026 · New paper release! Our follow-up to MAST AdaMAST: Adaptive Failure Taxonomies for Improving LLM Agents AdaMAST gives an agent system a memory of its own failures: a compact, validated taxonomy of named failure codes induced from its traces, reused by search, runtime monitoring, and trajectory selection. February 2026 · ADRS series Automating Algorithm Discovery: A Case Study in Improving Multi-Agent Reasoning Systems using MAST (Part 2) The evaluator is the bottleneck: with MAST-based feedback instead of binary reward, evolutionary search finds higher-performing multi-agent reasoning systems on Olympiad-level math. December 2025 · with IBM Research Why Do Enterprise Agents Fail? Insights from IT-Bench using MAST 310 SRE traces, three models, one lesson: success rates hide the failure signature. MAST separates fatal failures from benign friction and turns IT-Bench scores into an engineering roadmap. December 2025 · ADRS series Automating Algorithm Discovery: A Case Study in Improving Multi-Agent System Design using MAST Replacing hand-tuning with OpenEvolve: MAST failure modes as the fine-grained signal that lets evolution mutate a multi-agent architecture toward a 7x better failure rate. November 2025 · Tutorial How to Use MAST to Improve Your Agents A step-by-step tutorial: annotate a LangGraph agent's traces with the MAST Annotator, read the failure-mode histogram, and rewire the system for a 53% relative gain — same LLM.