OpenAI Expands Codex Into Autonomous Desktop Agent With Computer Control Capabilities

Apr 14, 2026

OpenAI has updated its Codex tool, transforming it from a coding assistant into an autonomous agent that can control a user's computer by seeing the screen and using a cursor. The update, which adds an in-app browser and image generation, is a direct challenge to competitor Anthropic's Claude. Internally, OpenAI already uses similar agents to autonomously manage its vast data infrastructure, signaling a strategic shift toward agentic AI.

Codex autonomous infrastructure agents

▪Internally, OpenAI uses Codex-powered AI agents to autonomously run its data platform, where they debug failures and manage releases.
▪One internal OpenAI agent manages updates for Apache Spark systems, while another acts as an "always-on on-call assistant" for new issues.
▪According to OpenAI, 80% of its own staff use Codex, highlighting its use by non-programmers.

OpenAI data platform scale

▪Event volumes on OpenAI's streaming systems have increased by approximately 50 times in one year.
▪OpenAI's internal data platform supports over 3,500 users and manages more than 600 petabytes of data across 70,000 datasets.

Agent data foundation requirements

▪Data agents can fail if the underlying data foundation has issues like missing metadata or siloed data across systems.
▪According to OpenAI's Emma Tang, data agents require a strong "data foundation," including a unified data lake, metadata, lineage, and permissions.

Agent operational workflow automation

▪The computer control capability is initially available only on macOS and will be released to EU and UK users at a later date.
▪OpenAI's updated Codex desktop app can operate a user's computer by seeing the screen, clicking, and typing with its own cursor.
▪New features in Codex include an in-app browser, image generation via gpt-image-1.5, and a "memory" to recall context from past sessions.
▪The new "background computer use" feature allows multiple agents to run in parallel on a Mac without interfering with the user's work.
▪OpenAI has added over 90 new plugins to Codex for integrations with tools like Atlassian, CircleCI, GitLab, and Microsoft Suite.
▪Codex can now schedule future tasks for itself and "wake up" to continue long-term projects across days or weeks.

OpenAI Anthropic competitive positioning

▪On the Terminal-Bench 2.0 benchmark, OpenAI's agent scores 77.3% compared to 65.4% for Anthropic's Claude.
▪The Codex update is a direct competitive response to Anthropic's Claude Code and Cowork products, which also feature computer control.
▪OpenAI is evolving the Codex app into a "Super App" intended for broader knowledge work beyond just software development.

Agent trust validation frameworks

▪To ensure human oversight, the system exposes an agent's assumptions, chain of thought, generated queries, and confidence levels for review.
▪OpenAI uses continuous evaluation frameworks to validate agent outputs against expected results and trusted "golden" sources.