Codex is replacing efficient primitives with slower AI workflows

I’ve been using Codex for real development workflows (game dev, frontend, UI systems, creative tooling), and I think there’s an important UX issue emerging with “agent-first” interfaces.

I understand the vision behind Codex:

• multi-agent workflows,

• orchestration,

• autonomous tasks,

• async execution,

• long-running jobs, etc.

I also understand why OpenAI may not want to build “just another VSCode fork”.

But right now, Codex often feels like it removed extremely efficient developer primitives before AI became reliable enough to replace them.

The result is that many workflows become slower and more frustrating than in a traditional IDE.

A few concrete examples:

1. Filesystem / Explorer

A classic filesystem explorer is not an outdated abstraction. It is an extremely efficient cognitive and visual tool.

Expanding: assets → audio → sfx

takes less than a second. No prompt. No latency. No interpretation. No AI cost. No ambiguity.

By comparison, asking an AI: “show me all sfx music files”

introduces:

• friction,

• latency,

• probabilistic interpretation,

• additional cognitive overhead.

AI should augment filesystem navigation, not replace it.

1. Direct manipulation vs AI interpretation

Sometimes I ask Codex to … change a color.

But in a traditional IDE, I can simply:

• click the color square next to a hex code,

• open a native color picker,

• visually explore the full color space in real time,

• instantly see what feels right.

This takes less than a second.

Using an AI for this is often worse:

• slower,

• more expensive,

• less precise,

• dependent on interpretation,

• and it breaks creative flow.

This is a fundamental distinction: direct manipulation vs probabilistic interpretation.

For many creative workflows, direct manipulation is superior.

1. Creative workflows need visibility

In game dev / UI / creative coding workflows, developers constantly:

• browse assets,

• inspect folders,

• compare files,

• preview images/audio,

• tweak values,

• use terminals,

• inspect diffs,

• navigate visually through project structure.

Removing these primitives compared to traditional IDEs like VSCode, Cursor, Windsurf or JetBrains products creates unnecessary friction in real-world workflows.

The filesystem is not just storage. It is part of how developers think. And access to the code is not just about writing everything manually. Sometimes it’s simply about quickly adjusting a value, tweaking a parameter, or changing a small detail directly.

My concern is that some current AI interfaces assume: “If AI can theoretically do something, it should replace the old tool.”

But many classic tools are already extremely optimized for human cognition:

• explorers,

• terminals,

• color pickers,

• timelines,

• node graphs,

• viewports,

• curve editors, etc.

AI is most powerful when layered on top of these tools, not when replacing them entirely with chat.

What I (and probably many advanced users) would love to see is:

• multi-agent orchestration,

• async tasks,

• AI-native workflows,

• real time collaboration,

BUT ALSO:

• integrated filesystem explorer (even a simple traditional sidebar with folders, subfolders, files, extensions, and expand/collapse arrows would already remove a huge amount of friction)

• embedded browser/devtools,

• lightweight document/media preview (not full Office-like applications, but simple built-in viewers/editors for common formats such as .txt, .md, .csv, .docx, .xlsx, images and audio files, to avoid constantly breaking workflow by switching between multiple external tools)

• direct visual editing,

• low-friction interaction.

The future probably isn’t: “replace all tools with chat”.

It’s: “deeply integrated human + AI collaboration.”

Right now, Codex sometimes feels like some incredibly efficient primitives were removed before the AI became good enough to compensate for their loss.

Thanks for taking the time to read this feedback.

(English is not my native language, so I used ChatGPT to help me write and structure this feedback more clearly.)