Build Hour: API & Codex

The paradigm for AI in software development is undergoing a significant transformation, moving from simple autocomplete and pair programming to a new era of agentic delegation. This evolution allows developers to delegate entire engineering tasks, from planning to execution, to AI systems. This shift is powered by increasingly capable models like GPT-5.4 and new tools designed to manage agentic workflows.

What’s New: Codex and API Capabilities

The Codex Desktop App

The Codex desktop app represents a move away from the IDE and toward a higher-level operational mode focused on managing and reviewing agents. Key features include:

• Agent Management: The app is designed to manage multiple agents across large, complex tasks and workflows.
• Cross-Platform Support: Now available on Windows with native sandboxing, meeting developers where they are.
• Skills and Apps: Users can provide agents with context and expertise through “Skills” and connect to external tools through “Apps” (formerly connectors).
• Work Trees: The app integrates git worktree functionality, creating sandboxed environments that allow agents (and developers) to work on multiple tasks in parallel within the same repository without conflicts.
• Automations: A powerful feature that allows users to run commands on a regular schedule. Use cases include reviewing open PRs for merge conflicts, summarizing recent git history for stand-ups, or managing Slack messages.

GPT-5.4 and API Enhancements

The release of GPT-5.4 brings state-of-the-art capabilities for agentic workflows, supported by several new API features:

• GPT-5.4: Features native computer-use capabilities, supports up to one million tokens of context, and introduces a tool_search tool. This enables “progressive disclosure,” where the agent can intelligently search and pull in only the necessary tools from a large set, rather than requiring all tools to be in the context from the start.
• Hosted Shell: Provides the model with a secure, containerized environment to execute bash commands, putting the power of coding agents directly into the API.
• Skills API: Allows developers to upload skills and reference them by an ID, making it easy for the model to access specialized knowledge.
• Websocket Mode: For extremely tool-heavy use cases, switching to websockets can improve latency by 20-30%.

Harness Engineering: Building for Agent Collaboration

“Harness Engineering” is a methodology developed from the experience of building a million-line codebase where 100% of the code was written by Codex. The core principle is to shift the engineer’s role from writing code to designing and managing a system that enables a team of agents to produce high-quality, reliable code.

The Core Philosophy: No Human-Written Code

By enforcing a strict “no humans write code” rule, the team was forced to solve problems at a systems level. The engineer’s job becomes encoding their knowledge, taste, and standards into the codebase itself, creating guardrails for the agents. This approach led to a significant increase in throughput, scaling from the equivalent of a quarter of an engineer to 3-10 engineers’ worth of output per person.

Combating “AI Slop” with Encoded Standards

A common concern with AI-generated code is “AI slop,” or code that developers don’t like. The solution is to articulate precisely why the code is undesirable and encode those standards directly into the repository.

• From Implicit to Explicit: If an agent produces code with a suboptimal pattern (e.g., re-implementing a concurrency helper instead of using a canonical, instrumented version), the solution is to create a lint rule that statically disallows that pattern.
• Encoding Non-Functional Requirements: Knowledge that often lives in conversations, Slack threads, or tribal knowledge (e.g., security best practices, reliability patterns like retries and timeouts) must be written down in documentation or encoded in reviewer agents.
• The Flywheel Effect: As each engineer on the team adds their unique expertise (e.g., frontend architecture, backend infrastructure) to the system’s guardrails, every agent in the system becomes more effective. This creates a flywheel where the collective knowledge continuously improves the quality and alignment of the AI-generated code.

Customer Spotlight: Basis’s Agent-First Engineering Culture

Basis, an agent platform for accountants, has deeply integrated these principles into its engineering and company culture. Their goal is to engineer the company to move at a speed that would be physically impossible to achieve through hiring alone.

Mindset Shift: From Doing to Managing

The most significant change is the mindset shift from being a “doer” to a “manager” of agents. This requires building a system where agents can work effectively and self-verify. Key elements include:

• Clear Standards and Context: Maintaining up-to-date standards, context, and processes is crucial. Basis uses agents.md files at the root and module levels to provide scoped instructions.
• Tooling for Agents: Internal development tooling must be accessible to agents via APIs so they can validate their own work, just as a human developer would.

Encoding Context Beyond the Codebase

Basis maintains two monorepos:

1. Arnold: The production codebase.
2. Atlas: A monorepo for all non-code company context, including operating principles, team structures, and strategic documents.

By putting all company context into a version-controlled, machine-readable format, agents have the information they need to help with higher-level tasks like planning and decision-making.

Tools for Agent Collaboration and Decision-Making

As code generation becomes a commodity, the main bottleneck shifts to decision-making. Basis has built internal tools to accelerate this:

• Paper: An internal application that renders specs, skills, and documentation from the codebase in a collaborative web interface. This allows teams to plan, comment, and resolve issues directly within the context of the repository.
• Dotnotes: A system where an agent records the key decisions made during a session in a .notes folder. This creates a historical “debugger” for commits, allowing future developers (or agents) to understand the reasoning behind a particular implementation.
• Sub-agents: Basis uses specialized sub-agents for tasks like enforcing standards or managing pull requests, which the main Codex agent can call upon.

• Tags:
• Developer Tools
• Feature
• AI Engineering
• Agentic Workflows
• Codex
• Harness Engineering
• GPT-5.4