8 Major Codex Updates in One Month: Evolving from Code Agent to Digital Colleague

From Backup to Main Player: Codex's Ambitious Transformation

When Codex CLI first launched last April, it was nothing more than a terminal Agent that could write code — no clear advantage over Claude Code, and its model capabilities were even slightly weaker. But a post published by OpenAI on April 16th completely redefined its positioning — titled "Codex for Almost Everything," signaling that Codex was no longer just a coding tool, but aimed to become an AI Agent capable of doing almost anything.

The 8 updates released this past month have transformed Codex from a code assistant into a "digital colleague." Let's break down each new capability.

Computer Use: The First Step Beyond the Terminal

Codex can now see your screen, click UI elements, and type on your computer with its own cursor. Once authorized, it can operate any app on your desktop.

Computer Use is a technical paradigm that lets AI directly control computer graphical interfaces. Under the hood, it typically relies on a "screenshot → visual understanding → action generation" loop: the model receives a screenshot, identifies UI element positions, outputs mouse coordinates and keyboard commands, then takes another screenshot to confirm results. Anthropic was first to commercialize this capability in October 2024, with OpenAI following suit. The core challenge lies in UI uncertainty — buttons shifting with window resizing, pop-ups interrupting workflows, loading delays causing misoperations — all requiring the model to have strong error recovery capabilities.

A typical scenario: having it extract color values from a Figma design file and write them into a Tailwind Config. It will open Figma on its own, take screenshots, perform image recognition, and write the configuration — all without you touching the mouse. This kind of cross-app work is very difficult for Claude Code to accomplish directly as a CLI Agent.

What's the trade-off? The permissions you're granting cover your entire computer, not just a specific project directory. The cognitive burden is an order of magnitude greater than with Claude Code.

Built-in Browser: Talk to the Agent by Pointing at Web Pages

Accompanying Computer Use is a built-in browser, and the most interesting design choice is: you can comment directly on web pages and feed those comments as instructions to the Agent.

For example: open a local preview page, select a button, and comment "this button state is too subtle, change it to a more prominent Primary Action" — done. This interaction pattern of pointing at a browser and talking to an AI Agent is far more intuitive than Claude Code's approach of connecting to apps via MCP — regular people don't need to understand what an API is.

MCP (Model Context Protocol) is a standardized protocol open-sourced by Anthropic in November 2024 that defines the communication interface between AI models and external tools and data sources. Its design philosophy is similar to USB-C: a unified interface specification that lets any model connect to any tool without developing a custom adapter for each combination. When Claude Code connects to external services via MCP, users need to configure the server side themselves — including installing dependencies, setting up authentication, and declaring tool lists. In contrast, Codex's plugin ecosystem follows an app store model, sacrificing flexibility for an out-of-the-box experience.

Automations: Giving the Agent Long-running Tasks

What's truly mind-bending is Automations — a persistent task system. You give the Agent a long-term goal, and it can reuse previous conversation context, automatically waking up on schedule to continue execution. OpenAI's exact words: "Potentially Across Days or Weeks."

For example, you could set: "Monitor this open-source project's issues daily. Whenever a performance-related one appears, compile the background, reproduction steps, and potential fix directions for me."