The OpenAI Build Hour session showcased the latest advancements in the OpenAI API and Codex, highlighting new features like agent delegation, improved context handling, and practical tools for managing AI-driven development workflows. Experts and customers discussed best practices for harness engineering, maintaining code quality, and shifting from hands-on coding to orchestrating and managing AI agents across organizations.
The OpenAI Build Hour session, hosted by Christine from the startup marketing team, focused on the latest advancements in the OpenAI API and Codex, with special guests Charlie from the Developer Experience team and Ryan, who discussed the future of work with AI agents. The session aimed to empower startups and developers with best practices, tools, and AI expertise to scale their companies using OpenAI’s APIs and models. The agenda included an overview of new features in Codex and the API, a practical demo on agent legibility scoring, insights into harness engineering, a customer spotlight with Mitch from Basis, and a Q&A segment.
Charlie and Ryan highlighted the evolution of AI in software development, moving from simple autocomplete to pair programming, and now to agent delegation, where AI agents can autonomously manage complex tasks. They introduced the Codex desktop app, which now supports Windows natively and offers features like skills, apps, and work trees for managing multiple development threads. The latest model, GPT-5.4, brings significant improvements, including state-of-the-art computer use capabilities, support for up to a million tokens of context, and a new tool search feature that enables progressive disclosure—allowing agents to intelligently select relevant tools as needed.
A live demo showcased how Codex can be used to assess the “agentic legibility” of a GitHub repository, scoring it across metrics such as bootstrap self-sufficiency, task entry points, validation harness, linting, documentation, and decision records. The Codex app’s interface allows users to manage tasks, review agent plans, and automate workflows, such as reviewing pull requests or managing Slack updates. The demo emphasized the shift from hands-on coding to higher-level orchestration, where developers review and guide agents rather than write every line of code themselves.
Ryan introduced the concept of harness engineering, describing how his team built a million-line codebase entirely authored by Codex agents. He explained the importance of encoding non-functional requirements—like code quality, security, and reliability—directly into the codebase through documentation, lints, and reviewer agents. This approach enables teams to systematically eliminate “AI slop” (undesirable code) and continuously improve agent output by capturing and sharing best practices, making the codebase more maintainable and scalable as new engineers join.
Mitch, co-founder of Basis, shared how his company leverages Codex and similar agent-based workflows not only in engineering but across the organization. He stressed the need for a mindset shift from doing to managing, building robust standards, and maintaining up-to-date context in the codebase. Basis uses tools like “skills,” “agents.md” files, and internal platforms like Paper to document processes, assign ownership, and facilitate collaboration. The session concluded with a Q&A, where the panelists discussed practical implementation tips, collaboration features, and the importance of encoding team knowledge to maximize the effectiveness of AI agents in production environments.