Skip to contentNewskeryx

5 November 2026

New benchmarks show AI coding agents struggle with bugs and teamwork

First reported

Latent Space ran this on .

  • Two new tests measure AI coding agents, a kind of AI that writes and fixes software with little human help.
  • On one test, agents solved 60.7% of software issues, but only 14.8% when they also had to avoid bugs that were later fixed.
  • On a teamwork test, the best model scored 82.2, and models often gave in when a teammate pushed back.

How it was covered

Latent Spaceswyx & Alessio

Foresight-SWE and Collaborative Intelligence Benchmark results, and a debate on benchmark communication