Building a Gemini Code Agent: Code Generation, Execution, and Debugging

Gemini's Built-In Code Execution

One of Gemini's most powerful features for agent development is native code execution. Instead of generating code and hoping it works, Gemini can write Python code, run it in a sandboxed environment, observe the output, and iterate if something goes wrong — all within a single API call.

This creates a true code agent loop: generate, execute, analyze, fix. The model does not just suggest code — it verifies that the code actually works.

Enabling Code Execution

Code execution is enabled as a tool, similar to function calling:

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

model = genai.GenerativeModel(
    "gemini-2.0-flash",
    tools="code_execution",
)

response = model.generate_content(
    "Calculate the first 20 Fibonacci numbers and show them as a formatted table."
)

print(response.text)

When code execution is enabled, Gemini writes Python code, executes it server-side in a sandboxed environment, and incorporates the actual output into its response. You see both the code and its real results.

Inspecting Execution Results

The response contains structured parts that separate code from output:

response = model.generate_content(
    "Generate a random dataset of 100 points and calculate basic statistics."
)

for part in response.candidates[0].content.parts:
    if part.text:
        print(f"TEXT: {part.text[:200]}")
    if part.executable_code:
        print(f"CODE:\n{part.executable_code.code}")
    if part.code_execution_result:
        print(f"OUTPUT:\n{part.code_execution_result.output}")
        print(f"OUTCOME: {part.code_execution_result.outcome}")

The outcome field tells you whether execution succeeded or failed. On failure, Gemini automatically attempts to fix the code and re-execute — you get the full debugging cycle in the response.

Building an Iterative Code Agent

Here is a code agent that handles complex multi-step programming tasks:

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

class CodeAgent:
    def __init__(self):
        self.model = genai.GenerativeModel(
            "gemini-2.0-flash",
            tools="code_execution",
            system_instruction="""You are a Python programming agent.
            When given a task:
            1. Break it into steps
            2. Write code for each step
            3. Execute and verify each step works
            4. If any step fails, debug and fix before continuing
            5. Always validate your final output""",
        )
        self.chat = self.model.start_chat()

    def solve(self, task: str) -> dict:
        response = self.chat.send_message(task)

        code_blocks = []
        outputs = []
        explanation = []

        for part in response.candidates[0].content.parts:
            if part.executable_code:
                code_blocks.append(part.executable_code.code)
            if part.code_execution_result:
                outputs.append({
                    "output": part.code_execution_result.output,
                    "success": part.code_execution_result.outcome.name == "OUTCOME_OK",
                })
            if part.text:
                explanation.append(part.text)

        return {
            "explanation": "\n".join(explanation),
            "code_blocks": code_blocks,
            "outputs": outputs,
            "all_succeeded": all(o["success"] for o in outputs),
        }

agent = CodeAgent()
result = agent.solve(
    "Read a CSV string with columns name, age, salary. "
    "Find the average salary by age group (20-29, 30-39, 40+). "
    "Format the results as a markdown table."
)

print(result["explanation"])
print(f"All code executed successfully: {result['all_succeeded']}")

Combining Code Execution with Function Calling

The code execution tool works alongside custom function calling. This lets the agent both run ad-hoc code and access external systems:

def query_database(sql: str) -> list:
    """Execute a SQL query against the analytics database.

    Args:
        sql: The SQL query to execute.
    """
    # In production, connect to your real database
    return [
        {"month": "Jan", "revenue": 150000},
        {"month": "Feb", "revenue": 175000},
        {"month": "Mar", "revenue": 162000},
    ]

model = genai.GenerativeModel(
    "gemini-2.0-flash",
    tools=["code_execution", query_database],
)

chat = model.start_chat(enable_automatic_function_calling=True)

response = chat.send_message(
    "Query the monthly revenue data, then calculate the trend line "
    "and predict next month's revenue using linear regression."
)

In this pattern, the agent calls your database function to get data, then uses code execution to run the statistical analysis. Each tool handles what it does best.

Test Generation Pattern

A practical application is generating and running tests for existing code:

agent = CodeAgent()

source_code = """
def parse_duration(text: str) -> int:
    parts = text.strip().split()
    total_seconds = 0
    i = 0
    while i < len(parts) - 1:
        value = int(parts[i])
        unit = parts[i + 1].lower().rstrip('s')
        if unit == 'hour':
            total_seconds += value * 3600
        elif unit == 'minute':
            total_seconds += value * 60
        elif unit == 'second':
            total_seconds += value
        i += 2
    return total_seconds
"""

result = agent.solve(
    f"Here is a Python function:\n\n{source_code}\n\n"
    "Write comprehensive unit tests for this function including edge cases. "
    "Execute all tests and report which pass and which fail."
)

FAQ

What Python libraries are available in the code execution sandbox?

The sandbox includes NumPy, Pandas, Matplotlib, and the Python standard library. It does not have network access or the ability to install additional packages. For tasks requiring other libraries, generate the code for local execution instead.

Is there a time limit for code execution?

Yes. Code execution has a timeout of approximately 30 seconds. Long-running computations will be terminated. Design your code agent prompts to break large tasks into smaller, faster steps.

Can the code execution sandbox access files I upload?

No. The code execution environment is separate from the Files API. If you need to process uploaded files with code, extract the content as text and pass it as part of the prompt, or use function calling to bridge the gap.

---

#GoogleGemini #CodeGeneration #AIAgents #Python #Debugging #AgenticAI #LearnAI #AIEngineering