From Cursor to Codex: One Developer's Real-World Take on the Shift from AI IDE to Super Agent

In April 2025, I began to clearly feel that AI in the coding domain was gradually moving toward commercial viability. Although it was still in the realm of assisted programming at the time, I subscribed to the then-popular Cursor at my own expense. $20 a month. Then, at the end of April 2025, my company started providing enterprise subscriptions for employees. Cursor even proactively refunded my unused money, which gave me a pretty good impression of Cursor at first.

I kept using it all the way until April 2026.

The overall experience was stunning. Compared to 'ancient programming,' Cursor represents a whole new era.

What Cursor does best is staying close to the coding scene.

Tab completion, Cmd+K, Agent, Plan Mode — all inside the editor. Modifying functions, completing pages, explaining code — all very smooth. Humans set the direction, and AI assists on the side. Of course, you can also let AI write the code in Cursor first, then switch to your familiar IDE for manual corrections. Additionally, Cursor lets you choose your preferred mainstream models; it's a third-party IDE, not tightly bound to any single model.

At that time, Cursor still had quite a few issues. For example, there was no directly usable Cursor plugin in Android Studio (AS). To keep code positions synced between Cursor and AS — so I wouldn't have to relocate code after switching back to AS — I had to install a plugin called IDE Sync. Another example: Cursor's CLI had plenty of pitfalls, and MCP and later SKILLS recognition initially had bugs.

Of course, none of these were real problems. AI develops too fast, and Cursor's version iterations are also very fast. After successive version improvements, combined with the continuous enhancement of model capabilities, starting from the second half of 2025 — especially as Claude surged — I clearly felt that Cursor's code understanding and coding ability had directly stepped up a level.

What impressed me most was this: before, for business details, I might have needed to ask a developer (honestly, our project's business logic is complex enough). But as Cursor iterated, it gradually became the case that I just asked Cursor directly. Later on, I wrote less and less code myself; Cursor could handle almost all scenarios.

On February 14, 2026, I published an article on my public account. It was entirely based on Cursor — within a short time, with zero manual coding, from design to implementation, I built a fully functional mini-program with decent visual effects, even though I had no prior experience with WeChat mini-program development.

At that moment, I felt that the end times for traditional programming were drawing near.

In April 2026, Cursor's pricing and enterprise policies became complicated. Cursor shifted from fixed licensing to usage-based billing, making costs higher and unpredictable. Based on this, the company decided that the China region would adopt OpenAI Codex to replace Cursor.

Just like the joke on Maimai: Who would have thought — before we lost our jobs, the company's Cursor lost its job first.

Starting mid-April 2026, I fully embraced Codex.

Codex gave me a very different impression from the start. Cursor had always felt more like an IDE — an IDE that accompanies code, just with AI capabilities on the side. But Codex feels like an Agent that can handle various work scenarios and tasks.

Of course, starting in April 2026, Cursor's new version also underwent a major redesign, continuously shifting from an IDE form to an Agent form, sparking widespread discussion in the community. Personally, I think this path and direction are correct. It's the general trend.

Codex's capabilities are truly strong. Coding, in its positioning, is probably just one of the core scenarios. Codex's update frequency is even more intense — recently, almost every two days there's a new version. Its interaction model directly reveals its ambition: to act as a super Agent, orchestrating capabilities across various scenarios.

After more than half a month of use, Codex is indeed easy to use — it can go end-to-end directly.

It feels a bit like the OpenClaw Agent approach, but it makes 'connections' much simpler and more user-friendly. It doesn't require users to do complex configurations; instead, through built-in 'plugins,' users simply enable the plugins they want to directly connect to external tools or applications.

At the same time, it can enable powerful features like Computer use. The newly released Codex Chrome plugin further strengthens the ability to automate browser operations, ultimately enabling full automation of many workflows. Even for complex tasks, Codex can handle them directly.

Of course, compared to Cursor, purely in terms of coding ability, overall, Cursor + Claude 4.7 still feels stronger than Codex + GPT-5.5, but the gap isn't huge — it's basically sufficient.

Another point worth mentioning: I wrote earlier about GPT Image 2 — its image capabilities are incredibly powerful, which gives Codex strong image capabilities directly in image-related scenarios.

Let me give a final summary: Cursor made me feel the shock of AI programming for the first time; Codex made me truly feel, for the first time, that AI is not just a programming assistant but a real Agent — more powerful and easier to use than previous Agent tools like Lobster. It's a true super Agent. Codex is clearly closer to the next-stage form.