OpenAI Codex: From Code Model to AI Agent Ecosystem

🧭 Context

This document captures research on OpenAI Codex as part of a broader study of popular AI code agents. The goal is to understand how production AI agents are designed and evolved — lessons to inform building a custom AI code agent.

Study path: Aider (articles, completed) → Codex (this document) → more agents to follow.

📛 Naming: What “Codex” Actually Means

“Codex” is an overloaded name at OpenAI. It has referred to three distinct things over time:

When people say “Codex” today, they almost always mean the agent platform, not the original 2021 model.

📅 Development History

The evolution of Codex tells the story of AI moving from autocomplete → agent → platform:

Era 1: The Model (2021–2023)

Era 2: The Agent is Born (2025)

Era 3: Full Platform (Late 2025–2026)

timeline
    title Codex Evolution
    section Model Era
        Aug 2021 : Codex model (GPT-3 fine-tune)
        2021-2022 : Powers GitHub Copilot
        Mar 2023 : Model deprecated
    section Agent Era
        Apr 2025 : CLI open-sourced (TypeScript)
        May 2025 : Cloud agent + codex-1 model
        Jun 2025 : Rust rewrite begins
        Aug 2025 : IDE extension
    section Platform Era
        Feb 2026 : Desktop app + GPT-5.3-Codex
        Mar 2026 : GPT-5.4 + Security agent
        Apr 2026 : Full suite mature

The Big Picture

2021  Completion    → predict the next token
2023  Conversation  → understand and respond to instructions
2025  Agency        → plan, use tools, execute autonomously
2026  Platform      → multi-model, multi-interface, multi-agent

Each step added a capability layer: completion → conversation → agency → platform.

🔀 The Four Interfaces: Comprehensive Comparison

Basics

Execution Model

Feature Matrix

Best For

🏗️ Architecture Highlights

Desktop App (Electron + Rust)

The architecture spans three process layers:

• Electron/React: windowing, ProseMirror editor, OAuth2 authentication
• Rust core: the actual intelligence — agent loop, sandboxing, tool execution
• 70-method IPC API surface between layers
• Built-in automation/cron system for scheduled background tasks

CLI (Rust)

• Originally TypeScript/Node → rewritten in Rust (95.6% as of Apr 2026)
• Reasons for rewrite: zero-dependency install, native security bindings, lower memory (no GC)
• Sandboxing modes: workspace-write (default, scoped) and danger-full-access
• MCP support for extensibility via ~/.codex/config.toml

Cloud

• Two-phase runtime: setup (network-enabled, installs deps) → agent (offline by default)
• Each task runs in isolated container with repo pre-loaded
• Container state cached up to 12 hours
• Secrets encrypted separately, available only during setup phase, removed before agent runs

🔑 Models Powering Codex

💡 Lessons for Building an AI Code Agent

From studying Codex’s evolution, several design principles emerge:

1. Start with CLI — Codex started as a simple terminal tool. CLI-first is the fastest way to iterate on an agent loop without UI complexity.

2. Tool use is the core — The agent’s power comes from its tools (file read/write, bash, git, web search), not just the model. Design the tool layer carefully.

3. Sandboxing matters early — Codex invested in sandboxing from day one. Letting an agent run code without guardrails is a non-starter for real usage.

4. Extensibility via protocol — MCP (Model Context Protocol) became the standard plugin mechanism. Building on open protocols beats proprietary plugin systems.

5. The CLI is the open-source part — For studying how agents actually work, the Codex CLI source code (Rust, Apache 2.0) is the primary learning resource.

6. Model-agent separation — The agent layer is model-aware but model-independent in design. The same agent framework ran on codex-1, GPT-5, GPT-5.2, GPT-5.3, and GPT-5.4 — swapping models without rewriting the agent.

🔗 References

• Codex CLI — GitHub
• Codex Developer Docs
• Codex App Features
• Codex IDE Extension
• Codex Cloud Environments
• Codex Sandboxing
• Codex Pricing
• Introducing Codex — OpenAI Blog
• Unrolling the Codex Agent Loop — OpenAI Blog
• Codex CLI Rust Rewrite — InfoQ
• Codex Architecture — Jiwei Yuan