OpenAI has significantly upgraded its Codex development system, granting it direct control over macOS applications and introducing a memory function. The move sharpens competition with rivals like Anthropic's Claude Code in the automated software engineering space.

Technically, the update allows Codex agents to operate desktop apps in the background without interrupting a user's primary workflow. Multiple agents can run in parallel. For engineering teams, this architecture could streamline front-end testing and interaction with legacy applications lacking APIs. The feature, launching initially for ChatGPT-linked macOS users, highlights a practical trade-off: rapid deployment on a single platform versus immediate cross-OS availability. EU rollout is delayed.

Beyond desktop control, the update bundles several system-level enhancements. Codex integrates new plugins for GitLab and Atlassian Rovo, adds native web browsing with page annotation, and incorporates an image generation model. A new scheduling capability lets Codex autonomously queue and resume long-running tasks, reducing manual orchestration overhead.

The most notable architectural shift is the introduction of an opt-in memory feature, currently in preview. This allows Codex to retain context—like user preferences or complex corrections—across sessions. The design aims to reduce repetitive instruction and improve output consistency, though it introduces new considerations for data persistence and privacy. This personalization will first reach Enterprise and Edu tiers.

Collectively, these updates represent OpenAI's strategic push to build a more autonomous, context-aware engineering agent, directly responding to feature advances in the competitive AI coding assistant market.

Source: The Verge