Developer says Codex breaks down on complex codebases and drifts off-plan
AlexanderWillard · reddit · 2026-07-26
A user says Codex becomes unusable once a codebase gets complex or long, even with additional agents reviewing the work.
- The model starts out well, but once bugs and complexity appear, it allegedly goes off the rails, invents details, and ignores instructions.
- It may also re-architect the app despite explicit guardrails.
- The author says they had a 6-hour session that still ended badly.
- In their setup, they prefer Fable and Opus for instruction following, with Sol used as planner and GLM 5.2 for code review before shipping.
The post is essentially a frustrated field report on how hard it is to keep coding agents on-plan in real projects.
More from coding & agent
- Agent Skills are loaded in one prompt, with caching reducing multi-turn token cost — dotey · 2026-07-26
- Grok Build Web UI adds a local-first dashboard for live CLI sessions — elonmusk · 2026-07-26
- LLM config tuning thread weighs KV cache tradeoffs in an 80K-context Laguna setup — kingo86 · 2026-07-26
- OpenAI rolls out ChatGPT Voice on desktop with voice control for multiple agents — steipete · 2026-07-26
- Agensis adds a collaborative workspace for human-agent projects on your own hardware — jasonkneen · 2026-07-26
- Codex backs up old high-school hard drives and resurfaces deeply awkward files — shakoistsLog · 2026-07-26