Hajra Naz
AI is transforming software development, shifting much of the routine work to swarms of agents and subagents. As developers explore new ways for humans and AI to collaborate, even the most advanced labs are struggling to keep up.
The current trend is agentic software development — systems in which AI agents operate autonomously on coding tasks. Tools like Claude Code and Cowork have epitomized this approach. Meanwhile, OpenAI has been steadily evolving its Codex tool, which launched as a command-line utility last April before expanding to a web interface a month later.
Now, OpenAI is taking a major leap forward. On Monday, the company unveiled a new macOS app for Codex, incorporating many of the agentic workflows that have gained traction over the past year.
The app is designed to run multiple agents in parallel, coordinate their skills, and integrate other advanced coding workflows. The launch comes just under two months after GPT-5.2-Codex, OpenAI’s most powerful coding model, debuted — a move aimed at enticing users of Claude Code and similar platforms.
“If you really want to tackle complex work, 5.2 is the strongest model by far,” CEO Sam Altman told reporters during a press call. “The challenge has been usability, so putting that level of capability into a more flexible interface is going to matter a lot.”
Benchmarks, however, paint a more nuanced picture. GPT-5.2 currently leads TerminalBench, a test measuring AI performance on command-line programming tasks. Yet Gemini 3 and Claude Opus agents have posted comparable scores — slightly lower, but within the margin of error.
Similar trends appear on SWE-bench, which tests AI’s ability to fix real-world software bugs, suggesting no definitive edge for GPT-5.2. Agentic workflows, in particular, remain difficult to benchmark, and user experience can vary widely across state-of-the-art models.
Read More: OpenAI Introduces Prism to Help Scientists Work Smarter With AI
The Codex macOS app introduces several features intended to match or surpass the capabilities of Claude-based competitors. Users can now schedule automations to run in the background, with results queued for review upon return. The app also allows developers to choose agent personalities, from pragmatic to empathetic, tailoring the AI to their workflow and preferences.
For OpenAI, the most compelling advantage remains speed of development. “You can start from a blank slate and build a sophisticated piece of software in a matter of hours,” Altman said. “As fast as I can type ideas, that’s the limit of what can be created.”
