google-adk-code-executor

name	google-adk-code-executor
description	ADK code execution backends. Use when agents need to write and run Python code — BuiltIn, Container, VertexAI, GKE, or local executors.

Google ADK — Code Executors

Overview

Code executors let agents write and execute Python code. ADK provides multiple backends with different isolation levels.

Executor	Isolation	Use Case	Install
BuiltInCodeExecutor	Gemini sandbox	Simple code tasks	google-adk
UnsafeLocalCodeExecutor	None (local process)	Development only	google-adk
ContainerCodeExecutor	Docker container	Self-hosted production	google-adk[extensions]
VertexAiCodeExecutor	Google Cloud sandbox	Managed production	google-adk[extensions]
GkeCodeExecutor	Kubernetes pod	GKE-based production	google-adk[extensions]
AgentEngineSandboxCodeExecutor	Agent Engine	Vertex AI Agent Engine	google-adk[extensions]

BuiltInCodeExecutor (Default)

Uses Gemini's built-in code execution capability. No additional setup.

from google.adk.agents import Agent
from google.adk.code_executors import BuiltInCodeExecutor

agent = Agent(
    name="coder",
    model="gemini-2.5-flash",
    instruction="Write and execute Python code to solve problems.",
    code_executor=BuiltInCodeExecutor(),
)

UnsafeLocalCodeExecutor (Development Only)

Executes code directly on the host machine. Never use in production.

from google.adk.code_executors import UnsafeLocalCodeExecutor

agent = Agent(
    name="dev_coder",
    model="gemini-2.5-flash",
    instruction="Write and run Python code locally.",
    code_executor=UnsafeLocalCodeExecutor(),
)

ContainerCodeExecutor (Docker)

Runs code inside a Docker container for isolation.

pip install google-adk[extensions]

from google.adk.code_executors import ContainerCodeExecutor

agent = Agent(
    name="sandboxed_coder",
    model="gemini-2.5-flash",
    instruction="Write and execute code safely.",
    code_executor=ContainerCodeExecutor(
        image="python:3.11-slim",
        packages=["pandas", "numpy"],
    ),
)

VertexAiCodeExecutor (Managed)

Google Cloud-managed sandbox. No container management needed.

pip install google-adk[extensions]

from google.adk.code_executors import VertexAiCodeExecutor

agent = Agent(
    name="cloud_coder",
    model="gemini-2.5-flash",
    instruction="Write and execute code in a managed sandbox.",
    code_executor=VertexAiCodeExecutor(
        project="my-gcp-project",
        location="us-central1",
    ),
)

GkeCodeExecutor (Kubernetes)

Executes code in a Kubernetes pod on GKE.

pip install google-adk[extensions]

from google.adk.code_executors import GkeCodeExecutor

agent = Agent(
    name="gke_coder",
    model="gemini-2.5-flash",
    instruction="Execute code on the cluster.",
    code_executor=GkeCodeExecutor(
        cluster_name="my-cluster",
        project="my-gcp-project",
        location="us-central1",
    ),
)

AgentEngineSandboxCodeExecutor

For agents deployed to Vertex AI Agent Engine:

from google.adk.code_executors import AgentEngineSandboxCodeExecutor

agent = Agent(
    name="engine_coder",
    model="gemini-2.5-flash",
    instruction="Execute code in the Agent Engine sandbox.",
    code_executor=AgentEngineSandboxCodeExecutor(),
)

Choosing an Executor

Requirement	Recommended Executor
Quick prototyping	BuiltInCodeExecutor
Local dev with full libraries	UnsafeLocalCodeExecutor
Self-hosted production	ContainerCodeExecutor
Google Cloud production	VertexAiCodeExecutor
GKE-based infrastructure	GkeCodeExecutor
Deployed to Agent Engine	AgentEngineSandboxCodeExecutor

Key Rules

• BuiltInCodeExecutor is the safest default — uses Gemini's built-in sandbox
• UnsafeLocalCodeExecutor has NO isolation — code runs with full host access
• Production deployments should use Container, VertexAI, or GKE executors
• Extension executors require pip install google-adk[extensions]
• Code executor is set on the agent, not the runner
• Only ONE code executor per agent (set via code_executor field)

Related Skills

• google-adk-llm-agent — Agent configuration (where code_executor is set)
• google-adk-deploy — Production deployment considerations