OpenAI Agents SDK 0.14.0 New Features and Sandbox Execution Environment

The latest update of the Agents SDK announced by OpenAI on April 15, 2026, marks a significant turning point in enterprise agent development. The new version 0.14.0 adds the ability to safely execute file operations, command execution, and code editing in a sandboxed environment.

According to the OpenAI team, “developers need not just the best models, but systems that allow agents to inspect files, execute commands, write code, and continue working over many steps.” This announcement provides a concrete solution to the constraints of traditional agent development. (Source: openai.com)

How the Sandbox Execution Environment Works

The new SandboxAgent class provides the functionality to run agents in a controlled workspace. The following code example demonstrates how to implement a data room analysis agent:

from agents import Runner
from agents.sandbox import Manifest, SandboxAgent, SandboxRunConfig
from agents.sandbox.entries import LocalDir
from agents.sandbox.sandboxes import UnixLocalSandboxClient

agent = SandboxAgent(
    name="Dataroom Analyst",
    model="gpt-5.4",
    instructions="Answer using only files in data/. Cite source filenames.",
    default_manifest=Manifest(entries={"data": LocalDir(src=dataroom)}),
)

result = await Runner.run(
    agent,
    "Compare FY2025 revenue, operating income, and operating cash flow with FY2024.",
    run_config=RunConfig(
        sandbox=SandboxRunConfig(client=UnixLocalSandboxClient()),
    ),
)

This implementation ensures that the agent only accesses files within the specified directory, preventing unauthorized access to external systems. The UnixLocalSandboxClient guarantees safe execution in a Unix environment. (Source: openai.com)

Harness and Compute Separated Architecture

The new SDK separates the harness (control layer) and compute (execution layer) to improve security, durability, and scalability. This design allows for independent management of the agent’s execution environment.

The harness is responsible for:

  • Managing the agent loop
  • Controlling tool invocation
  • Applying security policies

The compute layer is responsible for:

  • Actual code execution
  • File system access
  • System command execution

This separation enables developers to minimize security risks in production environments while providing agents with powerful execution capabilities. (Source: openai.com)

Installation Procedure and Pricing

The SDK can be installed using pip or npm:

pip install "openai-agents>=0.14.0"

In a TypeScript environment:

npm install openai-agents@latest

According to OpenAI’s announcement, the Python package has been downloaded approximately 14.7 million times in the past 30 days, and the TypeScript package has been downloaded approximately 1.5 million times. This usage scale indicates that the SDK is widely adopted in actual production environments.

As for pricing, the official announcement does not provide detailed information, but it is expected to follow the existing OpenAI API pricing structure. Developers can check the latest pricing information in the OpenAI developer documentation. (Source: developers.openai.com)

Summary

  • Run pip install "openai-agents>=0.14.0" to evaluate the new sandbox features
  • Use the SandboxAgent class to build secure agents with restricted file access
  • Configure UnixLocalSandboxClient to introduce a sandbox execution environment that meets production security requirements
  • Check the OpenAI official documentation for pricing details and enterprise features
  • Migrate existing agent implementations to the harness-compute separated architecture to improve scalability