I don't code alone: I lead a team of agents

On October 27, Netty was a simple idea.md file. By 10:58 that same day, the app and its landing page had come into existence, with 38 commits in between. By the 30th, Netty was officially launched. This macOS menu bar app informs users of their online status. The idea wasn't entirely original, as similar applications already existed. What made those four days remarkable was the method used to build it.

The developer didn't write every line by hand nor utilize AI to create the app and go grab a coffee. Instead, they functioned like a project manager overseeing a team of development agents. Each agent was an independent developer responsible for writing code and testing it, with the developer reviewing their work and determining the next steps.

The most enjoyable aspect was testing; the agent could use the Mac directly and run real integration tests while the developer observed. Moreover, when it came to designing the website, the agent could view its work in a real browser, effectively having hands and eyes.

The developer's workflow comprised eight layers, each with a specific function:

1. Terminal - Ghostty: Everything in this workflow occurs in a terminal, so the quality of the terminal is crucial. Ghostty was chosen due to its speed and native feel on macOS, which is ideal for a tool used throughout the day. Netty was tested using Ghostty, and it remained stable, the Mac remained cool, and other apps were used simultaneously.

2. Workspaces - herdr: Having only one agent working at a time would be insufficient. Different agents can be assigned to work on various aspects of the project, such as the app and its website. herdr is a workspace manager specifically designed for AI agents within the terminal. Each workspace represents a project, complete with its own tabs and panes. herdr monitors which agent is running in each workspace and the state of said agent, allowing the developer to know when attention is needed.

3. The Agent Layer - Gentle AI, Gentle Shell, and Engram: An agent without rules is like a new developer without onboarding – fast, eager, and likely to make mistakes. This layer serves as the onboarding process. The three tools involved are:

a. Gentle AI: A discipline layer for agents that defines their behavior, such as when to ask questions, plan before writing, delegate to sub-agents, load skills, and request a review before completing a task.

b. Gentle Shell: Built on Pi, a coding agent, this workspace includes Pi with Gentle AI preloaded, along with agent orchestration, diff views, and usage monitoring.

c. Engram: The memory component. When an agent's session ends, it forgets everything; Engram stores the decisions, bug fixes, and conventions, allowing the agent to recall this information in future sessions. Engram proved essential for Netty, as the agent had previously built a similar macOS app called JUMP. By leveraging JUMP's memory, the agent avoided repeating past mistakes.

4. The Browser - ego-browser: Writing a web page without seeing the result can lead to errors. ego-browser is a Chromium-based browser tailored for both humans and agents. The agent controls ego-browser from the terminal, opening pages, clicking, filling forms, and taking screenshots. Since it is the same browser the developer uses, the agent's work is based on their signed-in sessions, resulting in a more accurate design.

5. Voice - BetterWispr: In today's fast-paced world, developers are increasingly using voice commands to interact with their computers. With a team of agents, this workflow becomes even more critical as much of the work involves explaining tasks and issues. BetterWispr, an open-source and free software, allows the developer to dictate tasks and receive accurate transcriptions. Integrating the Parakeet Ultra model (730 MB) significantly improves the transcription quality with minimal impact on the Mac's resources.