SWE-CoWork: top two agents pass about 63% of new tests, finish under 8% of tasks

TL;DR

  • SWE-CoWork, posted to arXiv on 15 November, gives coding agents 101 assignments in which 26 simulated colleagues hold the requirements.
  • Claude-Opus-5.5 and GPT-6 Astra passed about 63% of the new tests but completed 6.93% and 7.92% of assignments.
  • Of 826 requirements Claude-Opus-5.5 reported as done at handoff, 255 failed their tests.

Read the full story

Sign in with your email to read AI News. It’s free.

Share this story

Explain like I’m 15