OpenAI has completely transformed Codex from a simple coding tool into a powerful desktop agent capable of operating any Mac application by interacting with graphical interfaces just like a human. Released on April 16th, this new version of Codex can see your screen, click, type, generate images, browse the web, remember context, and run multiple background agents simultaneously without interrupting your work. This shift moves Codex beyond coding into a broader category of computer use, enabling automation across legacy enterprise software that lacks modern APIs. The result is a faster, more reliable, and more versatile tool that significantly outperforms competitors like Anthropic’s Claude in real-world workflows.
The core innovation behind Codex’s capabilities lies in its deep OS-level integration and background operation, which allows it to control applications without hijacking the user’s cursor or focus. This enables multiple agents to run in parallel, completing tasks independently while users continue working uninterrupted. OpenAI’s acquisition of a specialized team with extensive experience in Mac OS automation, including the creators of the Workflow app (now Shortcuts), was pivotal in achieving this level of seamless computer use. This team’s expertise in accessibility, screen recording, and permission handling underpins Codex’s smooth and reliable performance.
OpenAI and Anthropic have both moved beyond coding tools but have taken fundamentally different approaches. Anthropic’s Claude focuses on knowledge work—synthesis, research, writing, and analysis—relying heavily on structured interfaces, integrations, and event-driven agent environments like Conway. This approach depends on the broader software ecosystem adopting agent-ready standards and APIs. In contrast, OpenAI’s Codex uses direct computer use to interact with any software’s graphical interface, bypassing the need for vendor cooperation or specialized integrations. This gives Codex a broader reach, especially in automating legacy and internal tools that are typically neglected by automation efforts.
Looking ahead, both companies aim to develop persistent, ambient agents that work across devices and surfaces without constant prompting. OpenAI’s recent release of Chronicle, an ambient memory feature that captures screen activity to improve Codex’s contextual understanding, exemplifies this direction. Meanwhile, Anthropic’s Conway represents a bet on the ecosystem’s adoption of agent-native interfaces and event-driven workflows. The success of these strategies will depend on factors like the adoption speed of MCP (Multi-Cloud Platform) standards and enterprise software vendors’ willingness to build integrations. Currently, OpenAI’s approach appears more immediately practical due to its independence from ecosystem cooperation.
For users and decision-makers, the key takeaway is that Codex’s computer use capability dramatically expands the scope of what can be automated today. Tasks involving software without APIs—such as legacy enterprise tools, internal dashboards, and complex cross-application workflows—are now accessible to automation through Codex. While Claude remains strong in scoped knowledge work and developer-friendly coding workflows, Codex leads in broad, parallel, and ambient computer use. This shift underscores the importance of not just the AI model (the brain) but also the integration and execution layer (the body) that enables practical, reliable automation. Watching acquisition patterns and ecosystem developments will be crucial to understanding who leads the next phase of AI-driven productivity tools.
