openai

OpenAI Codex Evolves Into a True Desktop Agent: Computer Use, Memory, and Background Automation

• 7 min read

OpenAI Codex Evolves Into a True Desktop Agent: Computer Use, Memory, and Background Automation

On April 16, 2026, OpenAI released its most significant Codex update since the developer tool's launch, transforming it from a coding assistant into a comprehensive computer automation platform. The update introduces computer use capabilities for macOS, persistent memory, background agent execution, and image generation—features that collectively position Codex as OpenAI's answer to Anthropic's Claude Code and signal the company's ambition to dominate the emerging agentic development ecosystem.

Beyond the Editor: Codex Takes Over Your Desktop

The headline feature is background computer use, enabling Codex to operate your Mac's desktop applications through its own cursor—seeing, clicking, and typing alongside your work. Unlike previous iterations confined to terminal and code editor environments, Codex now navigates the full macOS ecosystem, from design tools like Figma to project management apps like Jira.

Multiple agents can work in parallel without interfering with your primary workflow. For developers, this enables scenarios previously requiring manual intervention:

The feature launches initially for macOS, with OpenAI committing to EU and UK rollouts "soon." Windows support remains unconfirmed.

Memory: The Persistent Context Layer

Perhaps the most transformative addition is memory—a feature allowing Codex to retain useful context across sessions. Unlike stateless interactions that begin fresh with each conversation, Codex now remembers:

OpenAI describes this as achieving "a level of quality previously only possible through extensive custom instructions." The company emphasizes that memory is opt-in, requiring explicit user activation—a notable contrast to persistent memory implementations by competitors.

Memory in Practice: Real-World Scenarios

Consider a typical development workflow: A developer spends 20 minutes explaining their application's authentication architecture, database schema, and deployment constraints to Codex. Previously, this context would be lost with the next conversation. With memory enabled, Codex retains this understanding, allowing subsequent requests like "Add user profile editing" to automatically incorporate the existing auth system and database patterns without re-explanation.

Image Generation Meets Development

Codex now integrates gpt-image-1.5 for generating and iterating on visual content. Combined with its new computer-use capabilities, this enables workflows that span the visual and technical dimensions of product development:

The integration is particularly significant for frontend and game developers, who can now iterate on design and implementation in a unified workflow rather than switching between design tools, code editors, and AI assistants.

The Plugin Ecosystem Expands

OpenAI introduced over 90 new plugins extending Codex's reach across the development toolchain:

| Category | Integrations |

|----------|-------------|

| Project Management | Atlassian Rovo, JIRA, GitLab Issues |

| CI/CD | CircleCI, Render |

| Code Review | CodeRabbit |

| Productivity | Microsoft Suite, Slack, Notion |

| Infrastructure | Neon by Databricks |

| Media | Remotion (programmatic video) |

| Specialized | Superpowers framework |

These plugins follow the Model Context Protocol (MCP) specification, allowing Codex to gather context and take action across the entire development lifecycle. The integration with GitHub enables automatic PR review comment resolution, while multi-terminal tab support and SSH connections to remote devboxes round out the professional developer experience.

Automation: Agents That Schedule Themselves

Codex's automation capabilities now support self-scheduling agents capable of waking themselves to continue long-running tasks. This enables:

The system preserves context from previous conversation threads, allowing agents to build upon accumulated knowledge rather than starting from scratch. This addresses one of the most significant limitations in current AI assistant implementations—the inability to maintain coherent long-term projects.

Proactive Suggestions: The Anticipatory Agent

Beyond reactive assistance, Codex now proactively suggests work based on context analysis. By examining projects, connected plugins, and memory, Codex can:

This represents a shift from AI as tool to AI as collaborator—one that understands not just what you're doing, but what you might want to do next.

The Competitive Landscape: Codex vs. Claude Code

This release arrives as OpenAI's rivalry with Anthropic intensifies. Just days earlier, Anthropic launched Claude Opus 4.7 with benchmark-leading coding performance. Internal OpenAI communications leaked earlier this month revealed CEO Sam Altman describing the company as "on the wrong side of history" regarding developer tools, with plans to "materially change our product strategy" in response to Claude Code's success.

The Codex update appears to be the first materialization of that strategic pivot. Where Claude Code excels at deep reasoning and long-horizon software engineering tasks, Codex differentiates through:

| Dimension | Codex | Claude Code |

|-----------|-------|-------------|

| Scope | Full desktop automation | Code-centric development |

| Memory | Explicit opt-in persistence | Conversation history only |

| Plugins | Extensive third-party ecosystem | Primarily Anthropic-native |

| Scheduling | Self-scheduling agents | Manual initiation |

| Visual | Integrated image generation | High-resolution analysis |

| Workflow | Proactive suggestions | Reactive assistance |

Developer Adoption and Limitations

With over 3 million weekly active developers, Codex has significant distribution advantages. However, the update introduces several limitations:

OpenAI's approach to memory—making it opt-in rather than default-on—reflects growing industry awareness of privacy concerns and user agency in AI systems. This positions Codex as the more conservative choice for developers wary of AI systems accumulating personal context.

Architectural Implications

The Codex update reveals OpenAI's product strategy for developer tools:

This stands in contrast to Anthropic's strategy of building the most capable coding model possible while avoiding direct platform competition. OpenAI appears willing to sacrifice some raw capability for workflow comprehensiveness and user experience integration.

What This Means for Development Teams

Organizations evaluating AI coding tools now face a genuine choice:

The ideal setup may involve both: Codex for workflow automation and cross-application tasks, Claude Code for intensive software engineering work requiring sustained reasoning.

Looking Forward: The Path to AGI

OpenAI explicitly frames Codex's evolution within its broader AGI mission: "narrowing the gap between what people can imagine and what they can build." The company promises "much more to come soon," suggesting this update represents an early phase rather than a final vision.

The trajectory is clear: AI development tools are evolving from autocomplete suggestions to autonomous collaborators capable of managing complex, multi-day projects across the entire software stack. The question is no longer whether AI will transform software development, but how quickly developers can adapt to leverage these capabilities—and which platforms will define the standards of that transformation.

For the 3 million developers already using Codex, today's update offers a glimpse of a future where the boundary between human and AI contribution to software becomes increasingly fluid. For the rest of the industry, it sets a new baseline for what developer tools must offer to remain competitive.

The AI coding wars are escalating. Developers are the winners.

--