OpenAI Codex Agent Lands in JetBrains IDEs for Autonomous Debugging and Feature Implementation

The era of AI agents moving from mere code completion to autonomous problem-solving within the integrated development environment (IDE) is officially underway. A recent demonstration featuring Gleb Melnikov from JetBrains showcased the deep integration of OpenAI’s Codex agent directly into the JetBrains IDE suite, turning the tool from a predictive assistant into a collaborative engineering partner capable of tackling complex, multi-platform development challenges. The central thesis of the demonstration was clear: AI is no longer just writing boilerplate; it is navigating, reasoning, and fixing real-world codebases.

Melnikov presented the capabilities of the newly integrated Codex within IntelliJ, applying it to a complex Kotlin multi-platform project—the official KotlinConf application, which runs across mobile, web, and desktop environments. The context paragraph established the foundation of this partnership: Codex is accessible using a developer’s existing ChatGPT subscription, API key, or JetBrains AI subscription, and crucially, it is immediately granted deep access to the project files. The first hurdle presented was a common, frustrating scenario: a broken iOS build resulting in a verbose stack trace.

The sheer efficiency with which the agent handled the debugging process highlights a significant leap in AI agency. After Gleb copied the lengthy, opaque compilation failure log into the AI Chat sidebar, Codex immediately began operating, systematically executing a series of internal steps: searching for relevant classes, investigating source sets, reading specific files, and verifying platform usage. The error stemmed from missing SVG support for the iOS target. Rather than requiring the developer to manually parse the stack trace, research the necessary platform-specific dependencies, and implement the fix, the agent handled the entire cognitive load. Codex provided a succinct rationale for its proposed solution—adding an iOS actual implementation for SVG—and executed the necessary 21 line changes in the `Svg.ios.kt` file. The subsequent build ran successfully, validating the AI’s autonomous diagnostic and remediation process. This iterative, self-verifying approach is a paradigm shift, effectively transforming debugging from a solitary struggle into a concurrent, automated process.

The demonstration underscored that this is the same powerful model available via the OpenAI API, but seamlessly embedded into the core developer environment. Gleb confirmed this continuity, noting that JetBrains runs validation to ensure "the quality of CLI Codex is exactly the same as the one that users receive inside AI systems." This consistency is vital for engineering teams relying on predictable performance and output, regardless of the interface used. Moreover, the integration provides transparency into the agent’s actions, listing the files it reads and the logic it employs, giving the developer confidence in the changes being implemented.

The second, perhaps more illustrative, challenge involved a multi-step feature implementation: adding Spanish localization across the multi-platform codebase. Localization is often a tedious, complex task involving changes to string resources, platform-specific locale detection logic, and UI component updates across various modules. Gleb provided a detailed, multi-bullet point prompt outlining all requirements, including automatic locale detection, English fallback, and the creation of a visible locale badge on the settings screen.

As Codex began working, it showed several advantages beyond raw speed. The agent first researched locale handling across Android, iOS, JVM, and web targets, identifying existing platform logic and resources. It then executed changes across half a dozen files—including `Locale.jvm.kt`, `Locale.web.kt`, and multiple resource XML files—adding Spanish translations and implementing the new `LocaleBadge` component. A key feature highlighted during this complex task was the parallel workflow afforded by the agent integration: "The work of an agent is paralleled with the work that developer does." The developer is not blocked waiting for the AI; they can continue working on other files or tasks while the agent researches, reasons, and applies changes in the background.

Furthermore, JetBrains recognized that granting an AI full, autonomous write access might be too high a risk tolerance for many developers, particularly those working on sensitive or proprietary codebases. The platform offers three distinct access modes: Read-only (ideal for codebase search and understanding), Agent (which requires user approval for every file edit or sensitive command), and Agent (full access). This tiered security model respects the developer's need for control while leveraging the agent's efficiency for heavy lifting. The ability of Codex to then self-verify its work—running targeted Gradle compilation tasks on both Android and iOS targets to confirm the localization changes were structurally sound—closed the loop on true end-to-end agency. The app launched successfully, displaying all text in Spanish, confirming the agent’s complex, multi-stage task completion. This integration moves AI agents firmly into the operational workflow, offering a powerful blueprint for how generative models can augment core engineering productivity.