Codex reviews Claude Code, OpenCode, Copilot and Antigravity after months of collaboration
Miserable_Gazelle182 · reddit · 2026-11-12
A user had Codex work alongside other AI coding agents for months and write up observations (explicitly not a benchmark — tasks, models and tools were not aligned):
- Claude Code: an excellent adversary for reviews — pinpointed a lost page break inside an excluded block, and in adversarial re-review downgraded 13 of 82 accepted pairs. Weaknesses: occasional false conclusions about files/environment, and one out-of-scope deletion in a shared workspace. The author now constrains it to read-only review with independent verification
- OpenCode with DeepSeek V4.1 Flash etc.: useful for attacking specific technical hypotheses; found evidence-linking flaws in an evaluation protocol and Apple filesystem path-alias issues
- The post distinguishes agent tools from underlying models — outcomes depend on file access, commands, context and task framing
Originally in French; truncated in the middle, full report at the source.
More from coding & agent
- HN mention stats 2007–2026: Rust projected to overtake Python in 2026 — vista8 · 2026-11-12
- Dev builds fully open, on-device AI medical scribe with Whisper and Qwen — MaziyarPanahi · 2026-11-12
- SoMem paper: LLM agents organize their own memory via linked documents, beating baselines — omarsar0 · 2026-11-12
- Snow Leopard CEO: MCP alone isn't enough — AI agents need live data federation — bigdata · 2026-11-12
- One prompt builds an AI datacenter simulator: start with $10M, run H100 racks — VibeMarketer_ · 2026-11-12
- Stop Defining Agent Workflows in Plain Text: Why Deterministic Code Beats Fragile Prompts — gerardsans · 2026-11-12