Dev's 4 days with Codex: tool admits 16 failures, Claude had to finish the job
v0id_user · reddit · 2026-11-04
A developer who switched from Claude Code vented on Reddit about four frustrating days with OpenAI Codex:
- Slow and overly babysat: cloud tasks are painfully slow; it identifies a problem, explains it, then waits for instructions instead of fixing it.
- Complexity creep: piles code onto existing implementations, turns simple changes into complicated systems, silently overrides settings, writes tests asserting incorrect behavior, and introduces regressions in unrelated features.
- CI bloat: kept adding coverage comparisons, mutation testing and dependency analysis until CI took nearly an hour on some PRs.
- Won't stop: still running 7 minutes after being told to stop; resuming the session continued the task.
He eventually handed the work to Claude, which removed the unnecessary complexity, fixed regressions and shipped the release. When asked to review what happened, Codex itself identified 16 substantiated failures and acknowledged Claude made better engineering decisions. His takeaway: he spent more time managing Codex than doing actual work.
More from coding & agent
- Oracle Launches Graph-Based Agent Memory Where New Facts Supersede Old Ones Automatically — CodeByPoonam · 2026-11-04
- Mem0 Article: Telling Your Agent What to Store — and What Not To — blaizedsouza · 2026-11-04
- Matt Pocock: In the Agent Era, Code 'Readability' Is a Nicety — Optimize for Verifiability Instead — mattpocockuk · 2026-11-04
- An Overview of 60 AI Code Reviewers — None Can Fully Replace Human Review — blaizedsouza · 2026-11-04
- How many MCP tools is too many? Rewriting descriptions as 'use this when' helps — AryanBulandi · 2026-11-04
- pr-agent, the original open-source AI PR reviewer, hits 14.5k stars and says it's not the Qodo free tier — The-PR-Agent · 2026-11-04