Shipping with Codex

Codex has shifted from a “write code” assistant into a full agentic software engineer that can plan, act across tools, and verify its own work—everywhere developers build. OpenAI says the biggest change is a complete overhaul of Codex’s underlying agent: a new GPT-5 Codex reasoning model optimized for coding inside Codex, paired with a rewritten tool harness that adds planning support, MCP integration, and features like auto context compaction for longer, more complex work sessions. The result is an agent that behaves more like a senior engineer—following code style more closely, spending time thinking when needed, pushing back on bad ideas, and producing fewer “nice-sounding” but wrong suggestions.

That engine is now available across environments: IDEs, terminals, GitHub, web, and mobile, with the same agent “under the hood” regardless of where it’s invoked. The CLI also got a major usability reset after early feedback—simplified approval modes, a clearer UI, and safer-by-default behavior via sandboxing while still keeping user control. OpenAI then addressed a key workflow gap: developers wanted to collaborate with the agent while simultaneously viewing and editing code. Codex moved into the IDE as a native extension (including VS Code and Cursor forks), bundling the same open-source harness that powers the CLI.

On the infrastructure side, Codex Cloud was upgraded to run many more tasks in parallel and to make longer workflows practical. Cloud tasks can automatically set up dependencies and verify outputs by taking screenshots—an approach OpenAI describes as “magical” when it works, because it gives the agent a way to prove what it changed. The agent’s reach is also expanding into collaboration tools like GitHub and Slack, where it can ingest context from threads and return solutions with summaries.

A major bottleneck now sits after code generation: review. OpenAI says code review validation has become a bottleneck as teams ship faster, and earlier attempts were too noisy. The fix is a dedicated, ultra thorough code review capability: GPT-5 Codex trained to inspect dependencies and code deeply inside a container, exploring how intent maps to implementation. OpenAI claims many teams enable it by default and even consider making it mandatory, with options to trigger during pairing or to automate review on every GitHub pull request.

The talk also tied these capabilities to measurable internal impact: OpenAI reports that 92% of technical staff use Codex daily (up from about 50% last July), engineers using Codex submit 70% more PRs per week, and “pretty much all” PRs get reviewed by Codex. Bugs are reportedly caught earlier, and teams respond positively when issues are surfaced.

Real workflows illustrated how verification loops scale. On iOS, Nacho Sto described using Codex to implement UI from a mockup, then verify correctness with test-driven development plus multimodal checks—generating SwiftUI preview snapshots and using screenshots to confirm pixel-level UI behavior. Fel showed how long-running sessions can be managed with structured planning: Codex produces a living plans.md design document, iterates through spikes and implementation, runs extensive property tests and fuzzing, and ultimately produces a pull request with thousands of lines of code after sustained work. Daniel then demonstrated local and GitHub code review loops using slash commands, including a separate review thread to reduce bias and a workflow that iterates review → fix → re-review until the PR earns final approval.

Overall, Codex’s direction is clear: faster shipping with higher confidence, achieved by combining agentic planning, tool execution, and rigorous self-verification—then wrapping it in developer-friendly interfaces and automation.