# Build Hour: API & Codex Source: https://www.youtube.com/watch?v=rhsSqr0jdFw Recap page: https://rapidrecap.app/video/rhsSqr0jdFw Generated: 2026-03-10T18:33:00.776+00:00 --- ## Quick Overview OpenAI's Build Hour session introduced the third phase of AI in software development: agent delegation, highlighted by new features in Codex including GPT-5.4 with KUA capabilities, up to a million tokens context, and tool search, alongside the Windows-native Codex desktop app featuring skills and apps for advanced workflow automation. **Key Points:** - The progression of AI in software development is now in the third phase, agent delegation, moving beyond autocomplete (phase one) and pair programming (phase two), enabling management of multiple agents for large tasks via the Codex desktop app. - New features include GPT-5.4, which offers native state-of-the-art computer use agent (KUA) capabilities, supports up to a million tokens of context, and uses a new tool search for progressive disclosure of tools. - The Codex desktop app is now available on Windows with native sandboxing and introduced features like 'skills' (context for agents) and 'apps' (formerly connectors) for integrating daily tools. - API updates feature a new code mode allowing the model to generate JavaScript for ripple execution, and a 'hosted shell' tool providing a container environment for executing bash commands securely. - Harness engineering focuses on making repositories 'agent legible' by encoding non-functional requirements through tests, lints, and documentation, exemplified by the Symfony repo analysis which received a B-grade on legibility metrics. - Basis co-founder Mitch demonstrated using Codex to automate company processes beyond coding, emphasizing the need to shift from doing to managing, and using 'agents.mmd' and 'skills' with owners for canonical context management. - Through harness engineering patterns, throughput increased from 'a quarter or half an engineer's worth' to 'three to 10 engineers worth of throughput per engineer' by encoding team expertise into the codebase constraints. **Context:** This OpenAI Build Hour session, hosted by Christine, Charlie (Dev X team), and Ryan (visiting from Seattle), focused on maximizing engineering work using OpenAI's Codex, specifically detailing recent updates to the Codex API and the desktop application. The session aimed to teach attendees how to use Codex for next-level engineering tasks beyond simple pair programming, featuring a demo on agent legibility scoring and a discussion on harness engineering principles. ## Detailed Analysis The session outlined a significant evolution in software development driven by AI, transitioning into agent delegation enabled by major updates like GPT-5.4. GPT-5.4 introduces KUA (computer use agent) functionality, supports a massive one million token context window, and implements 'progressive disclosure' via a new tool search to manage complex toolsets efficiently. The Codex desktop app, now on Windows, centralizes agent management using skills and apps to automate workflows, with speakers noting the app has mostly replaced the IDE for them. A demo illustrated 'agent legibility score,' which analyzes repositories against metrics like bootstrap self-sufficiency and validation harness to optimize them for agentic workflows, exemplified by analyzing the Symfony repository. Ryan introduced 'harness engineering,' a philosophy focused on encoding non-functional requirements (like linting standards, reliability patterns) into the codebase so agents reliably produce high-quality, mergeable code, citing a project where Codex wrote one million lines of code under strict constraints. Customer spotlight Mitch from Basis reinforced these concepts, explaining how Basis uses Codex to scale their agent platform for accountants by treating company context (operating principles, specs) as code in a separate repository ('Atlas') and utilizing sub-agents for specialized tasks like standards enforcement, ultimately achieving high velocity by making decision-making, rather than coding, the primary bottleneck. ### New Codex & API Features - Third phase of AI development is agent delegation - GPT-5.4 supports 1M tokens and KUA capabilities - Hosted shell provides containerized execution environment for bash commands - Tool search enables progressive disclosure of hundreds of tools ### Codex Desktop App Capabilities - App is now Windows native with secure sandboxing - Features include 'skills' for context injection and 'apps' for tool connection - Automations allow scheduled tasks like PR review and Slack management ### Agent Legibility Demo - Demo involved creating an app to score a GitHub repository on seven metrics for agentic legibility - Agent explored previous runs, demonstrating episodic memory - Work trees enable cheap, parallel work streams within the app ### Harness Engineering Principles - Goal is to remove humans from the loop by ensuring reliable agent output - Encoding non-functional requirements (e.g., linting rules, security practices) bans 'AI slop' from the codebase - Increased throughput from 0.5x to 3-10x engineer output by investing in building blocks and context ### Customer Spotlight - Basis - Basis uses Codex to engineer scaling beyond physical hiring limits - Mindset shift required from 'doing' to 'managing' agents - They use 'agents.md' files and 'skills' with defined owners for canonical context management ### Context and Decision Making - Decision-making becomes the bottleneck at high velocity - Basis uses 'dotnotes' to record agent decisions historically, acting as a debugger - They encode company context (operating principles) in a separate repo ('Atlas') for agents to use in planning.