Implement the agent loop

---

title: "Build Your Own Redis, Git & SQLite From Scratch" source: "https://app.codecrafters.io/courses/claude-code/stages/ff2" author: published: created: 2026-05-31 description: "Advanced programming challenges for experienced engineers. Rebuild real production tools like Redis, Git, and SQLite from scratch in your IDE." tags: - "clippings"

---

Implement the agent loop #ff2

In-progress

Upgrade

Stage #ff2

In-progress

Your Task¶

In-progress

Medium

In this stage, you'll implement an agent loop.

The Agent Loop¶

So far, your program handles a single interaction: send a prompt to the LLM, get a response, execute one tool if requested, and exit.

This works for simple tasks, but falls short when the task requires multiple steps (e.g., "read a file and fix any bugs").

For this stage, you'll implement an agent loop that repeatedly sends messages to the model and handles tool calls as needed, until the final result is received.

Here's the structure in pseudocode:

messages = [{ role: "user", content: prompt }]

loop:
    response = call_api(messages)
    append response message to messages

    if response has no tool_calls:
        print response.content
        exit

    for each tool_call in response.tool_calls:
        result = execute_tool(tool_call)
        append {
            role: "tool",
            tool_call_id: tool_call.id,
            content: result
        } to messages

Let's walk through each part of this loop:

1. Initialize the conversation: You already have an initial conversation history: the messages array with the user's prompt. Now you need to store this array so it can persist across iterations, since the loop will continuously append new messages to it: js [ { "role": "user", "content": "Summarize the README for me." } ]
2. Enter the loop: Start the loop with the same API request you already have (sending your messages and tool specifications to the model). The difference is that this request now sits in a loop, allowing it to run multiple times.
3. Record the assistant's response: Whatever message the model returns, add it to your messages array. If the model wants to use a tool, the response will contain a tool_calls array: js { "role": "assistant", "content": null, "tool_calls": [ { "id": "call_abc123", "type": "function", "function": { "name": "Read", "arguments": "{\"file_path\": \"README.md\"}" } } ] }
4. Execute tool calls: Check the model's response to see if it's requesting to use any tools. If tool calls are present:
   • a. Execute each requested tool (but do not print their result to stdout).
     • b. Add each tool call result to your messages array. Every tool result must:
     • Have the role field set to "tool" - Reference the corresponding tool_call_id - Include the tool call result as its content js { "role": "tool", "tool_call_id": "call_abc123", "content": "# My Project\n\nChemical expiry period: 6 months" }
5. Repeat until complete: Continue the loop until the model responds without requesting any tools (when tool_calls is missing or empty). At this point, print the final message content to stdout and exit.

Tests¶

The tester will create a Python project with:

• README.md
• Two Python files in app/ with randomized names

The tester will then execute your program like this:

$ ./your_program.sh -p "Use README.md to determine the chemical expiry period in months. Number only."
<expiry period in months>

The tester will verify that:

• The output is the correct expiry period
• Your program exits with exit code 0

Notes¶

• Do not print raw file contents when executing the Read tool. This differs from the behavior in the “Execute the read tool” stage.
• You can print intermediate debugging information to stderr. Print to stdout only when the final result is ready to be printed.
• You can also use finish_reason: "stop" from the first response choice as a signal to stop the loop.

How to pass this stage¶

1

Write code to pass this stage

Head over to your editor / IDE and implement your solution.

If you want a quick look at what functions to use or how to structure your code, we recommend looking at Code Examples.

2

Submit code

To run tests, make changes to your code and run the submit command:

codecrafters submit # Visit https://codecrafters.io/cli to install

Test Runner:

Ready to run tests

Hints¶

Filter by Python

Write

Preview

Python LEADERBOARD

Complete this stage to hit #10569

Ready to run tests

Show logs