Ampersand’s multi-agent mode scores 47/51 (92.2%) on SWE-bench Pro v2 HARD-51 versus base GLM-5.3’s 43/51 in the matched harness, humans& reports. Access is through its Agent API.

Official humans& coding replay interface showing three Ampersand agents coordinating commands and messages along a shared timeline.

Published 5 Nov 2026 · 22:36 UTC

humans& — Co-Coding announcement ↗ · Source dated 5 Nov 2026

Browse all Live updates · Read Features