엽니다AI 사본

Codex can spin up multiple independent agent instances simultaneously, each working in its own sandboxed git worktree. This means a developer can assign a bug fix, a refactor, and a new feature build all at the same time, reviewing the proposed changes as pull requests. The result is a multiplied throughput without any context collision between tasks.

The April 2026 update brought full background computer use to the macOS desktop app, allowing Codex to see, click, and type using its own cursor without disrupting the developer's active screen. An integrated in-app browser lets developers comment directly on live pages to give precise frontend instructions to the agent, making it especially valuable for UI iteration, game dev, and localhost app testing.

Codex connects directly to Slack, enabling any team member to delegate coding tasks or ask codebase questions from within a channel thread, exactly as they would with a human colleague. This removes the friction of switching between communication and development tools, and it is available to all ChatGPT Plus, Pro, Business, Edu, and Enterprise plans.

Codex now retains memory from previous sessions, storing personal preferences, corrections, and project-specific context to improve quality on future tasks without requiring repeated custom instructions. Automations can be scheduled so Codex wakes itself up to continue long-running work across days or weeks, covering everything from landing open PRs to staying on top of Jira threads and 개념 문서.

The GPT-5.5 model now powers Codex by default, achieving 82.7% on Terminal-Bench 2.0 and leading competing agents in 에이전트 코딩 and long-running professional work. A 1M token context window (available on GPT-5.4+) means large codebases with hundreds of interconnected files can be reasoned over in a single task without chunking or loss of coherence.

Codex integrates natively with over 90 plugins spanning tools like Atlassian Rovo (Jira), CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, Neon by Databricks, and Render. These give the agent context it needs from across your entire toolchain, not just the repo, allowing it to act on real engineering signals rather than abstract prompts.