AI agents often repeat failed strategies instead of using corrective feedback

Across four cooperative puzzle tasks, instructions targeting observed collaboration failures raised GPT-4o pairs’ average success from 60.5% to 70.5%.

Big Tech

Javier Hernandez · Andrew D. Wilson

Microsoft Research

Research Digest·12 November 2026·2 min read

Thread:Agent Self-Improvement

Illustration: AI agents often repeat failed strategies instead of using corrective feedback

In one line

GPT-4o agent pairs succeed in 60.5% of asymmetric puzzle conversations; a taxonomy-guided intervention raises success to 70.5%.

Before you read this

Three ideas this paper builds on

multi-agent coordinationThe collaborative operation of multiple AI agents to solve tasks more effectively or with greater capability.87 papers →agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real290 papers →failure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.140 papers →

Read it as

(LLMs) are AI programs that generate text. They are now used as autonomous agents that work together, but how well they collaborate is unclear.

To study this, researchers built a platform with puzzles where pairs of LLMs each see different information and must chat to complete the task.

On average, the agents succeeded in 60.5% of attempts. When the researchers added prompts to address frequent mistakes, the success rate rose to 70.5%. This shows that systematic analysis and targeted fixes can improve LLM collaboration, which matters as these systems take on more complex joint tasks.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

multi-agent coordination agentic evaluation failure attribution large language model