AI Coding: "No One Is Left Behind" – Key Principles of Agent Coding for Developers

The AI coding ecosystem contains numerous technical terms, such as 'harness' and 'context engineering,' whose meanings can be difficult to grasp. Since these terms were unfamiliar just a few years ago, developers can easily feel pressured to know all of them. This illustrates the practical difficulties developers face in understanding and keeping up with new concepts in a rapidly changing technological environment.

The speaker expressed vague fears about AI coding and the difficulty of getting started, mentioning the anxiety of feeling that it is complex yet also that one will fall behind if they do not begin. The fear many developers feel about being left out of this new trend is a common phenomenon, acting as one of the entry barriers to AI coding.

Theodor René Carlsen emphasized that no one has been left behind in the field of AI coding. He reassured developers by explaining that even if LinkedIn or a new boss claims developers are falling behind, it's merely a repetition of the same argument about constantly having to catch up with the latest technology. This message offers great comfort to developers who feel anxious amidst the rapid changes in AI technology.

AI coding tools are primarily composed of three important elements. First is the 'model,' which acts as the most crucial brain, accounting for 95% of the overall functionality. Adding any other features is useless if the model does not work correctly. Second is the 'harness,' and third is the 'agent.' Understanding these three elements is a prerequisite for effectively utilizing AI coding tools.

The speaker emphasized that the 'harness' is a core tool that helps the model perform actual actions. While the model merely receives text input and outputs text, the harness makes the model take specific actions. He explained the harness as a concept similar to cloud code, acting as an intermediary that connects the model's instructions to actual tasks.

When most people talk about an agent, it encompasses both the model and the harness. The core components of an agent are the 'system prompt' and 'tools.' The system prompt is the initial text passed to the model, defining its role and identity. For example, it can be used to set the model's personality, such as 'You are the best developer in the world,' and can be configured by the user.

Another core component of an agent is 'tools.' The speaker mentioned that he initially found the concept of tools difficult to understand when first using AI-assisted coding tools. The toolset provides agents with the ability to read files, execute terminal commands, and retrieve data from the web. This is an important element that enables agents to perform various tasks in a real development environment, beyond their simple text processing capabilities.

The speaker used the OpenAI JavaScript SDK as an example to explain the specific method of declaring tools. He stated that a function capable of reading files can be declared as a tool, and its description and parameter types can be specified. However, he pointed out that the implementation details of the actual function are not explicitly stated, implying the relationship between tool declaration and actual operation. This method is used to define what tasks an agent can perform and to provide those tasks in a format the model can understand.

The feedback loop for collaborating with AI is changing, moving towards granting agents more control. Many developers initially used cloud code or agents in VS Code like ChatGPT, asking questions and copy-pasting code. However, this method does not fully utilize the agent's potential. The trend is gradually shifting towards users relinquishing control to the agent rather than occupying a large part of the loop, owing to the agent's high intelligence. Even the speaker, who was initially skeptical, now acknowledges the AI tool's autonomous working capabilities.

AI agents possess the ability to autonomously achieve goals even when given ambiguous tasks. Agents can navigate codebases, read files, interpret their output, and act appropriately. They can even perform iterative processes such as running tests and modifying tasks based on the results. Users can control how much they intervene in this loop; some aim for full automation, while others participate more by steering the agent.

One of the core principles of AI agent coding is to transform all problems into 'text.' Since models are trained on text and output text, any problem that can be converted into a text format can be effectively processed by an agent. This is crucial for abstracting complex problems into a simple form that AI can understand and handle.

The speaker cited his experience wanting diagrams, especially dependency graphs, to understand a codebase. Instead of creating diagrams manually with tools like Excalidraw or Figma, he wanted an agent to do it. To achieve this, he utilized a markup language that could generate diagrams, ultimately enabling on-the-fly diagram generation from the codebase. This case demonstrates the practical application of the 'text task' principle: converting complex visual information into a text-based markup language for an agent to process.

'Context engineering' might sound like a new concept, but it is actually about what text you input into the model. The speaker advised treating the agent like a new colleague. Agents know how to write code and operate computers, but they might lack detailed knowledge about a specific domain or project. Therefore, it is important to provide clear and detailed context, just as you would brief a new colleague on the project's background and domain knowledge.

AI agents can sometimes go out of control or proceed in unexpected directions. The key method to steer the agent in such cases is to provide 'more context-related text material.' This means giving the agent additional instructions or constraints to guide it back to the desired path.

The 'agents.md' file is typically used to store project-specific information that the model does not know inherently. This file can contain highly specialized content for the codebase, such as instructions to boot PostgreSQL before running tests, or a glossary of terms for a specific domain language. This serves as crucial context, helping the agent understand the unique requirements of the project and perform tasks accurately.

The agents.md file should not include general operational information about TypeScript or PostgreSQL. The speaker compared AI to a very smart new employee who needs time to become familiar with a new environment. This means that instead of general technical knowledge that the agent already knows or can easily learn, only specific information unique to a particular project or domain should be selectively provided. This is a method to increase efficiency without overloading the agent's context with unnecessary information.

Skills can be useful for new libraries that the model hasn't yet grasped or for domain-specific content like Danish tax law. However, skills should be kept concise and added minimally only when there is a strong need to steer the agent extensively. This is because an excessive number of skills can hinder the agent's autonomy and efficiency.

It is important not to overlook that AI models are very proficient at using the command line. Therefore, instead of using a specific tool like GitHub MCP (Microsoft Cloud Protector) to open a pull request, utilizing basic command-line tools such as GitHub CLI (Command Line Interface) might be a method that the model understands much better and can work with more efficiently. This implies that leveraging the model's fundamental strengths is more effective than building complex, separate toolchains.

The speaker strongly asserted his hope that cloud-based closed-code environments are not the future. Such environments could allow vendors to monopolize harnesses and models, causing developers to lose control over the ecosystem. He emphasized that while a black-box model might be acceptable, the harness, which is within our control, should not be a black box, and warned against vendor lock-in. This is an important message for maintaining the autonomy and independence of the developer ecosystem.

Ultimately, maintaining open source will benefit the entire agent coding ecosystem, the speaker argued. He stressed that since AI coding tools run on developers' machines, developers must strive not to lose control over them. This means that preserving openness and developer autonomy is essential for fostering a healthy ecosystem in the long term, amidst technological advancements.