AI code Agents

AI is a reality, whatever we like or not. And it will transform the future of professional software development as we know it. We must embrace they change, learn how to use it, adapting our processes to take advantage.

What is a AI code agent?

First thing to clarify is that code agents are more than a chatbot, they are a AI-powered tool using LLM (large-language models) that can perform complex development task, like:

  • Repository analysis: analyze the codebase, identify dependencies, generate documentation and create diagrams.
  • Automatic debugging and fixing: identify bugs, performance issues, security vulnerabilities and automatically fix them.
  • Analyze code: identify code smells, suggest improvements, enforce coding standards and patterns.
  • Generate new features: given a feature request, generate the code to implement it.
  • Analyze test coverage: identify untested code, generate tests automatically, review current test cases and suggest new test cases.

There are two types of code agents:

  • Copilot-like agents: designed to assist developers in writing code, providing suggestions, completing code snippets, generating code based on natural language descriptions (like User Stories).
  • Autonomous agents: designed to perform complex development tasks autonomously without human intervention, such as planning, writing code, writing test, refactoring, debugging and more.

Devin AI, Claude Code, Cursor, GitHub Copilot are examples of code agents.

Key concepts

Let's review some concepts that are important to understand how code agents work:

  • models: code agents use LLM models to generate code. You can choose the model you want to use, depending on your needs and budget.
  • tool calling: is when the code agent executes commands, to know more of the context to know exactly what to do.
  • system prompt: used to set the role of the agent, giving instructions and defining the context.
  • user prompt: the specific task to be performed.
  • skills: by default, code agents know how to write code. But they can be extended to enhance their capabilities. Example: a series of checklist when reviewing the code.
  • context window: agent short memory. It contains all the information that the agent can use to perform the task. The more context you can provide, the better results you will get. But it is limited, so once you reach the limit, agent will start forgetting the oldest information.
  • RAG (Retrieval Augmented Generation): is a technique that allows the agent to retrieve relevant information from external sources (like documentation, codebase, etc) to generate more accurate and relevant responses.
  • orchestrator: inside the agent, this component analyzes the task and directs the flow of execution, deciding when to call tools, when to retrieve information, when to generate code, etc.
  • workflows: step by step process to perform a task. Example: to build a new feature, you can define the following workflow:
    1. analyze the user story
    2. plan the implementation, breaking down the user story into smaller tasks
    3. generate code for each task
    4. generate tests for each task
    5. review the code and tests
    6. commit and push the code to the repository
    7. deploy the code.
  • MCP (Model-Connector-Protocol): is a standard protocol that allows any AI agent to connect to other tools (github, slack...), avoiding building custom connector for each tool. Tools that implements MCP that can be used by any agent that support MCP.

How a code agent works?

Whenever you give a code agent a prompt, it first analyze the task and gather context using tools. Then, actions will be taken and finally results is verified. The process is iterative, agent can execute multiple iterations until the task is completed. The orchestrator will decide when to call the tools and when to generate code. While models will do the reasoning and generation, tools can do other tasks to gather more context and execute actions (like search files, online, do file operations...).

References