OpenAI launches macOS app for agentic coding

OpenAI has launched the Codex app for macOS, designed to enable developers to manage multiple AI agents simultaneously and collaborate on long-running tasks.

The app allows users to run agents in parallel, switch between tasks seamlessly, and review or manually edit agent outputs within the interface.

It supports worktrees, enabling multiple agents to work on the same repository without conflicts and maintain isolated copies of code.

The app integrates with existing Codex CLI and IDE extensions, allowing for immediate use with current projects.

Additionally, Codex has expanded its capabilities with “skills,” which enable the AI to perform tasks beyond code generation, such as information gathering and problem-solving.

OpenAI demonstrated the app by creating a racing game using various skills and tools, with the AI acting as designer, developer, and tester.

The new app aims to support more complex, multi-agent workflows for software development.

🔗 Source: OpenAI

🧠 Food for thought

Implications, context, and why it matters.

### OpenAI’s new app enters a crowded market for AI developers

OpenAI’s app arrives with momentum, yet Microsoft’s GitHub Copilot still leads enterprise coding, and 65% of enterprises prefer incumbent solutions when available 1.
Anthropic holds a stronger position in software development and data analysis among enterprises, and enterprise penetration rose 25% since May 2025, with 44% of enterprises using Anthropic in production 1.
Competition stays scattered across workflow tools, including Cursor for an AI-native editor experience and Aider for developers who prefer open-source, command-line interfaces 2.
GitHub Copilot also supports models from multiple AI providers, which fits the 81% of enterprises using three or more model families in testing or production 1.

### The new bottleneck is managing AI, not building it

Managing multiple agents shifts the productivity constraint toward human attention rather than AI capability 1.
Software developers move from writing code to directing specialized AI agents, as in a demo where OpenAI asked Codex to build a racing game from one prompt while it acted as designer, game developer, and QA (quality assurance) tester 1.
OpenAI’s long-term plan ties agent-based coding to automating knowledge work, rooted in the premise that code is the universal language for controlling computers 1.
This model adds risks that include reliance on autonomous systems and weaker foundational engineering skills when developers hand off refactoring and bug-fixing to agents 1.