agent-evals/RESULTS.md

Description

Agentic Tools

• pi
• opencode
• claude-code

Grading

Purely opinion based, but based on

• One-shot performance
• Follow up fixes
• UI design

Rankings

Overall

1. eval/pi-kimi-k2.5
2. eval/pi-qwen3-coder-next-80b
3. eval/pi-glm4.7
4. eval/pi-glm4.7-flash

Parameters: < 100B (Local)

1. eval/pi-qwen3-coder-next-80b
   • UI: 8/10
   • Fixes:
     • New File
     • Markdown Styling
2. eval/pi-glm4.7-flash
   • UI: ?/10
   • Fixes:
     • FE wouldnt run on first try
     • Had to downgrade a bunch of deps so it would run
     • Used legacy packages
3. eval/pi-devstral-small-2
   • UI: 2/10
   • Fixes:
     • ?

Parameters: > 100B (Hosted)

1. eval/pi-kimi-k2.5
   • UI: 9/10
   • Fixes:
     • Files List
     • Editor Scrolling
     • Duplicate .md Name
2. eval/pi-glm4.7
   • UI: 7/10
   • Fixes:
     • Files List
     • Markdown Styling
     • Delete Failure