Codex 4.0 Workspace Agent Upgrades

Workspace Context and Agent Persistence

OpenAI has shifted the Codex platform toward a persistent workspace agent model, moving beyond simple chat-based coding. The new Appshots feature for macOS allows users to capture the frontmost application window, including screenshots and text, by pressing both command keys. This provides the agent with immediate visual context for UI debugging or troubleshooting without requiring manual file uploads. Additionally, Goal Mode is now a stable feature across the app, IDE extension, and CLI, allowing the agent to maintain intent over long-running tasks. The system now includes improved progress tracking and stop-conditions to prevent token waste when the agent encounters blockers or usage limits.

Remote Execution and Team Infrastructure

Codex now supports remote computer use that persists even when the host Mac is locked, utilizing short-lived authorization tokens and manual unlock fallbacks to maintain security. For enterprise environments, plugin sharing is now available for ChatGPT Business, enabling teams to distribute standardized skills, MCP servers, and lifecycle hooks. CLI updates in version 0.133 and 0.134 have improved plugin discovery, marketplace-aware listing, and connector tool reliability, specifically allowing read-only MCP tools to execute concurrently.

Frontend and Browser Automation

Browser-based workflows have been refined through advanced in-app annotations, allowing users to point at specific UI elements to request changes in font size, spacing, or color. The browser agent now extracts image assets and structured data more efficiently via a read-only JavaScript sandbox. To improve user experience, the Chrome extension no longer creates excessive tab groups, opting instead for tab icons to indicate agent status. These reliability fixes aim to reduce the flakiness previously associated with browser-based agent tasks.