Qodo’s Codex plugin brings company coding rules and independent review into developers’ coding sessions to help catch bugs spanning multiple repositories, OpenAI said in a customer account published Nov. 2.
OpenAI supplies the models Qodo uses. It described the workflow in “Qodo builds context into AI code review with Codex.” The account provided customer examples and executive comments, without a measured improvement in bug detection or a defined study period.
Qodo co-founder and Chief Executive Itamar Friedman said customer environments can span 10,000 or more repositories, the separate stores that hold software code and its history. A change that appears correct locally can break a service elsewhere, he said.
The plugin gives Codex access to company engineering knowledge, coding rules and a separate review process through Qodo’s Agentic Toolbox, OpenAI said. That context can help connect dependencies across repositories, including those hosted by different providers or linked through Kafka, a system for passing data between services.
Before writing code, Codex can retrieve relevant rules, earlier proposed code changes and patterns through the plugin, the account said. Qodo then reviews the new changes and returns findings within the same coding session, before the developer submits them for review.
Nnenna Ndukwe, Qodo’s AI Developer Relations Lead, said developers use the workflow to build features and address technical debt, or accumulated maintenance work. She said they automate the handoff between Codex’s coding work and Qodo’s review using GPT-6 models.
Friedman said coding agents and review agents need different instructions and context, and each role may involve multiple agents. He described a proposed workflow in which agents discuss findings and send unresolved disagreements to developers for human review.
Developers would act as planners and reviewers in that arrangement, Friedman said. They could inspect unresolved issues together with the agents’ discussions, supporting context and earlier developer decisions. He presented that approach as a way to build trust in automated development.
Friedman also described a possible use for OpenAI’s Decisions API, a software interface, to choose which rules apply at different development stages. In his example, it could select relevant checks from 200 rules before a language model considers them. He said that could reduce selection time and cost while improving review quality.
Friedman described the longer-term aim as a software factory customers could assemble and own, combining coding and review agents with knowledge of production software. In that proposed system, company knowledge would guide code creation, separate agents would check changes, and people would resolve outstanding questions.
