After a quarter of growth, both the model and ecosystem have reached new heights.
Looking back, the piece I wrote last September, Everyone Can Code: Vibe Coding and Problem Scale, remains accurate.
Of course, during these three months, I’ve been even more into vibe coding than before. Now, let’s talk about some new insights.
No Silver Bullet
Most problems don’t have a standard solution—software engineering is all about trade-offs.
AI already possesses the domain knowledge needed for making decisions, but it still can’t make decisions for you. Even if every relevant factor is fed into its context, not everything can be quantified (and AI doesn’t necessarily help quantify them correctly).
However, AI can offer options and explain their rationale, making decision-making much easier.
The Doom of Junior Engineers
Pure junior programmers will become like textile workers—meaning they’ll no longer exist in the future.
You might ask, if there are no junior engineers, how does one grow into a senior? Don’t worry; we won’t need senior programmers either. Just as modern textile factories don’t need senior textile workers, their engineers didn’t come from sewing lines.
Although junior programmers will disappear, a new paradigm will emerge. Those who can master it will still be in demand. The core skill will shift from translating requirements into code to defining acceptance criteria and deep problem correction.
Of course, a factory won’t have many senior technicians—certainly not as many as the former textile workers.
A Heavy Sword Has No Edge; True Skill Appears Effortless
After using Codex (gpt‑5.2‑high, without the ‑codex suffix) and Claude Code (Opus 4.5) side by side, I found that a truly strong base model stands out.
Even though Claude Code crushes Codex in features and experience, Codex manages to catch bugs that Claude Code misses, and it often spots root causes that Claude Code overlooks.
In a duel where both models review the same piece of code, Codex almost always finds flaws in Claude Code’s review, while Claude Code gives Codex’s review a flawless pass.
It’s the same in daily development—Claude Code tends to be lazy and superficial, while Codex thinks several layers deeper.
The only three drawbacks of Codex: too slow, poor front‑end/UI domain knowledge, and bad aesthetics.
Debugging Becomes Interrogation
When code grows complex enough—and is AI‑generated—I often treat Codex like a suspect. As I review diffs, I ask: “Are you sure you handled concurrency here?” / “What happens if the API returns null?” / “Explain why you used this library instead of that one.”
And so, I’ve shifted from mechanic to prosecutor.
Rather than fixing issues directly, I now press the AI through intense logical questioning, forcing it to expose gaps in reasoning. Perhaps that’s why slow‑thinking models like Codex are irreplaceable in reviews—they can withstand interrogation better.
Beyond Coding
Ever since Claude Code was released, I’ve used it for tasks like organizing computer files—reading PDF contents, then batch‑renaming files based on what’s inside.
It felt a bit like a low‑budget Jarvis 😄 after all, this world too is a kind of runtime.
Later, I realized that Anthropic’s Agent SDK depends on Claude Code as its runtime. Is that paradoxical? Not necessarily.
So I suspect that the Coding Agent might actually be the early form of a General Agent, already validated by Anthropic—and worth exploring:
Code will shift from being a product of human logic to becoming middleware that controls the world.
We’ll no longer write programs for computers to execute, but write requirements for AI to generate code that acts upon the world.
AI‑Oriented vs Human‑Oriented
Now that Agents are first‑class citizens, the endpoint of software delivery is no longer just the human user—it’s also the agent user.
Previously, product managers cared about polished UI and smooth interaction. But those design choices—made to reduce human cognitive load—are now obstacles for Agents.
After all, when your user is an AI that can process tens of thousands of tokens per second, it doesn’t need CSS; it needs precise JSON and deterministic logic.
We may soon need a whole range of services for Agents.