GLM-5.3 Shows a Different Chain-of-Thought Personality Under Claude Code vs Codex Scaffolds
repligate · x · 2026-11-01
Tim Hua observed that GLM-5.3's chain of thought changes noticeably depending on the agent scaffold: it behaves one way under the Claude Code scaffold and another under the codex scaffold — like having a different personality with different friend groups. The observation, amplified by repligate, suggests model behavior is shaped substantially by the scaffolding and prompts around it, not just the weights.
Related event: GLM-5.3 Chain-of-Thought Shifts Personality Across Agent Scaffolds(2 posts)→
More from coding & agent
- Pre-call guardrails that stop a runaway AI agent before the bill arrives — MaverikSh · 2026-11-01
- Browser-Based Video Mapping Built with Claude Code, Open for Agent Remixing — ismaelfaro · 2026-11-01
- Dev runs Codex-driven Blender 3D modeling agent through 8 iterations, hits prompt limits — ZeYanjie · 2026-11-01
- Craft team open-sources Craft Anything for understanding how any software works — ismaelgokufox · 2026-11-01
- "Tell AI to build Netflix": viral exchange reignites debate on AI coding limits — burny_tech · 2026-11-01
- Engineer observes emergent behavior: coding agent writes its own tasks and asks AX to run them — rakyll · 2026-11-01