GLM-5.3 shows strikingly different chain-of-thought under Claude Code vs Codex scaffolds
sebkrier · x · 2026-11-01
- Tim Hua compared GLM-5.3's chain of thought on the same design-review task under the Claude Code vs Codex scaffolds, and the difference is striking.
- Under Claude Code, the thinking is long and structured: it enumerates three work items, evaluates invariants one by one, and flags test blind spots (e.g. no canary test for the fast path against real-world bytes).
- Under Codex, the reasoning summary is more fragmented: it deliberates how to phrase sandbox-caused test failures, then lists positive code-organization notes and abstraction suggestions (typing.Protocol decoupling, splitting SessionRepository/SearchService, etc.).
- Takeaway: model behavior isn't determined by weights alone — agent scaffolding and prompts visibly reshape reasoning style and focus.
More from coding & agent
- Karpathy reveals his coding setup: Claude Code desktop app, Opus 5.5 at medium effort — Yuchenj_UW · 2026-11-01
- Qwen Code Desktop v0.26.5 ships background agent tracking and split-view shortcuts — github-actions[bot] · 2026-11-01
- Indie hacker's viral video made with a Skill that auto-builds social-proof walls — yihui_indie · 2026-11-01
- Inside Deepgram's Builder's Lab: 45 devs live-building voice agents, legal questions raised — schwentker · 2026-11-01
- Vapi CEO: 'Promise something impossible, then deliver it' — voice agent insights — schwentker · 2026-11-01
- Voice agent design wisdom: sometimes you need to add friction, and context selection is the real work — schwentker · 2026-11-01