OpenAI's Codex in ChatGPT: AI for Engineering Teams

In a significant step towards integrating artificial intelligence deeper into software development workflows, OpenAI has showcased Codex in ChatGPT, a powerful tool designed to assist engineering teams. The video, titled "ChatGPT Work for Engineering Teams," illustrates how Codex can streamline various engineering tasks, from routine bug fixes to complex refactoring and migrations, all while keeping human engineers in the driver's seat.

This demonstration from OpenAI highlights a future where AI acts as an intelligent co-pilot, taking on the repetitive and time-consuming aspects of software development. By automating the grunt work, Codex aims to free up engineers to focus on more creative and strategic challenges, ultimately leading to faster development cycles and higher quality code.

StartupHub data

In our directory

Slack

Channel-based messaging and collaboration platform for team communication and workflow automation

Founded

2013

Location

San Francisco, United States

Funding

$1.4B

OpenAI

OpenAI is an AI research and deployment company dedicated to ensuring that artificial general intelligence benefits all of humanity.

Founded

2015

Location

San Francisco, United States

Valuation

Private / $100B+ est

Linear

Fast, elegant issue tracking and project management system for software teams and AI agents

Founded

2019

Location

San Francisco, United States

Funding

$134M

The full discussion can be found on OpenAI Youtube's YouTube channel.

Engineering with Codex in ChatGPT

The core promise of Codex in ChatGPT is to take engineering work from an initial issue to a tested, review-ready code state. The video opens by posing a question: "What if engineers could spend less time fixing bugs and more time shipping better software?" This sets the stage for how Codex aims to transform the engineering workflow.

The demonstration begins with a user interacting with ChatGPT Codex, asking it to investigate a specific bug. The prompt is direct: "Can you check @Linear for the seat selection bug and figure out a plan to fix it?" This immediately highlights Codex's ability to integrate with external issue tracking systems, in this case, Linear.

Automated Bug Investigation and Planning

Upon receiving the request, Codex springs into action. The interface shows the AI's thought process, detailing its steps: using a "Linear skill" to find the actual issue, reproducing the bug in a running application, tracing relevant Swift/UI code, and then proposing a fix. This multi-step process mimics a human engineer's approach to debugging, but at an accelerated pace.

The AI reports back on its findings, stating it reproduced the bug by selecting specific options in a simulated app, noting that a call-to-action (CTA) remained disabled. It identifies that the seat map uses a local selection while the CTA derives completeness from a different, stale state source. This detailed analysis demonstrates a deep understanding of the codebase and the underlying issue.

Following the investigation, Codex presents a comprehensive "Fix AENG-976: Seat Selection State" plan. This plan includes a summary of the bug, a detailed implementation strategy involving modifying the BookingStore to be the single source of truth, removing view-local state, passing selections through the store, and deriving CTA text and enabled state from the booking state. Crucially, the plan also includes a validation section, outlining steps to run package tests, rebuild and relaunch the app, and verify the two-traveler flow, ensuring the fix works as intended and preserves existing functionality.

Building, Validating, and Sharing the Fix

After presenting the plan, Codex prompts the user with "Implement this plan?" with options to submit or provide feedback. Once approved, the AI proceeds to build and validate the fix. The interface shows progress as Codex implements the changes, confirming the state-ownership guidance and creating new test files.

The video then visually demonstrates the fix in action on a simulated iPhone. The seat selection bug is gone, and the application behaves as expected, with the CTA correctly reflecting the chosen seat and advancing to traveler details. This real-time validation is a powerful feature, allowing engineers to see the impact of the AI's changes immediately.

Finally, Codex handles the crucial step of sharing the review-ready code. The user requests, "Can you draft the PR and share it for review?" Codex leverages a GitHub publish workflow to create a focused commit, push it, and open a draft pull request (PR). It then uses a Slack outgoing message workflow to share the verified review link in the designated channel. The output clearly lists the PR, a Slack review post, the branch, commit ID, and validation status, including the number of tests passed and simulator flow verification.

Impact on Engineering Workflows

OpenAI's vision for Codex in ChatGPT is clear: to empower engineers by automating tedious tasks, allowing them to focus on higher-value work. By handling the entire lifecycle of an engineering task from issue identification to a review-ready PR, Codex promises to significantly accelerate development cycles and improve code quality. The control remains with the engineer, who can review, modify, or reject the AI's proposed changes, ensuring that human oversight is maintained throughout the process. This approach could redefine how engineering teams operate, fostering greater efficiency and innovation.