Qodo Brings Business Context to Codex for Review

Qodo combines its own independent review tools with OpenAI's Codex code generator. The company aims to integrate internal rules, change history, and inter-repository dependencies into the IDE, leading to broader automation and a sorting of checks via OpenAI's Decisions API.

Qodo combines artificial intelligence and software engineering

Qodo merges artificial intelligence methods, traditional software engineering approaches, and scripting. The company's internal knowledge guides the coding agents, an independent review examines their suggestions, and human intervention resolves any outstanding questions. Itamar Friedman presents this goal as a "software factory" that clients can compose and own, bringing together coding and review agents while maintaining a view of the software's evolution in production. Alongside OpenAI, Qodo is preparing for a future where agents rely more on the company's experience and developers reserve their expertise for critical points.

Qodo wants to leverage OpenAI's Decisions API to select rules

Itamar Friedman identifies an opportunity in OpenAI's recently launched Decisions API to determine, at each stage of development, which rules and agent skills apply, rather than querying a language model for every decision. He gives the example of a scenario where Decisions would select which of the 200 rules to check before a language model analyzes their relevance. According to him, this sorting could speed up and reduce the cost of selection while improving the quality of the review. He believes that the Decisions API can enhance trust in the process.

A Codex plugin that connects internal rules, dependencies, and independent review

Qodo utilizes OpenAI's models, including GPT-6 Astra, to centralize engineering knowledge, conduct code reviews, and ensure software quality tracking. Its Codex extension inserts this context and verification into the development environment. It is designed for client environments that may have 10,000 repositories or more, where a local change can affect a downstream service. This plugin allows Codex to access the knowledge base, organizational rules, and independent verification through the Agentic Toolbox. It takes into account dependencies between different repositories, including those managed by various Git providers or connected via Kafka. Qodo establishes these links so that Codex can understand the system as a whole and identify major bugs.

Automating the loop before the PR: retrieve, modify, control

Before Codex produces code, the Qodo plugin can collect relevant rules as well as the history of pull requests. After making modifications, Qodo evaluates them and then sends its analyses to Codex within the same session, even before the developer submits a pull request. Nnenna Ndukwe notes that some teams use this process to deliver features and mitigate technical debt, relying on the GPT-6 models integrated into Codex, which allows for automating the transition between code generation and independent review. She believes these models are extremely capable.

"Yin and yang" agents and disagreement-centered escalation

Itamar Friedman highlights a pair of agents with complementary roles: one generates code and the other critiques it. Each of these functions requires distinct configurations and contexts and can mobilize multiple agents. The idea is for these agents to confront their results and only pass on disagreements to humans, while developers handle planning and review, having access to the exchanges between agents, context, and decision history. He sees this as a way to build trust, and Nnenna Ndukwe believes this operating mode makes the whole process more efficient and values expertise.

Internal uses: GPT-6 Astra in Extra High and Ultra modes

Internally, Nnenna Ndukwe uses the Extra High and Ultra effort settings of GPT-6 Astra, both in automated development workflows and in her knowledge activities. She programs Codex to provide a weekly summary, highlight areas for improvement, browse conversations on agentic engineering on social networks, and suggest content ideas for developers. For her own code changes, she entrusts it with analyzing and suggesting agent skills to create or refine.