Skip to content

Latest commit

 

History

History
286 lines (205 loc) · 11.2 KB

env_service.md

File metadata and controls

286 lines (205 loc) · 11.2 KB

EnvService Overview

To evolve agents in a data-driven manner within dynamic environments, a robust and scalable environment management layer is essential. EnvService fulfills this role by providing a unified service that manages the lifecycle of training environment instances, enabling seamless interaction, evaluation, and control during agent training.

EnvService is designed to operate independently of other components and can be launched and used in isolation. It exposes a well-structured API and integrates with distributed execution frameworks like Ray, enabling support for large-scale agent evolution across heterogeneous environments.

Core Responsibilities of EnvService

  • Dynamic Environment Instantiation
    Dynamically create environment instances based on task specifications, parameters, and environment types (e.g., appworld, bfcl,webshop).

  • Lifecycle and Resource Management
    Track environment usage and automatically release inactive instances to ensure efficient resource utilization.

  • Unified Interaction Interface
    Provide a RESTful API for interacting with environments — including initialization, stepping, evaluation, and instance cleanup.

  • Modular Environment Loading
    Support dynamic import and registration of custom environment modules without modifying core service logic.

  • Asynchronous and Distributed Execution
    Leverage Ray actors to run each environment instance independently, enabling scalable and concurrent execution.

EnvService serves as the execution backbone for the agent evolution pipeline, enabling consistent interaction patterns, efficient resource handling, and flexible integration with various synthetic or user-defined environments.

Environment Setup

We provide a simple setup script and a Dockerfile to help you get started quickly.
In this subsection, we explain the environment installation process.
The launch instructions are provided in the next section.

More environments will be supported soon.

Appworld

To install the appworld environment, navigate to the following directory and run the setup script. This script will configure environment variables, create a Conda environment with Python 3.11, install necessary dependencies, initialize Appworld, and download required data:

cd env_service/environments/appworld
bash setup.sh

BFCL

To install the BFCL environment, navigate to the following directory and run the setup script. This script will configure environment variables, create a Conda environment with Python 3.11, install necessary dependencies, initialize BFCL, and download required data:

cd env_service/environments/bfcl
bash setup.sh

Launch Environment Service

After installation, you can launch Environment Service with the following commands:

Appworld

To launch the appworld environment service, run the script in the env_service/launch_script/appworld.sh:

source $(conda info --base)/etc/profile.d/conda.sh
conda activate appworld
cd env_service/launch_script
bash appworld.sh

By default, the service will start on 127.0.0.1:8080.
If you need to change the host (--portal) or port (--port), edit the following line inside env_service/launch_script/appworld.sh:

exec python -m env_service.env_service --env appworld --portal 127.0.0.1 --port 8080

Feel free to replace the IP address or port number as needed to suit your environment.

BFCL

To launch the bfcl environment service, run the script in the env_service/launch_script/bfcl.sh:

source $(conda info --base)/etc/profile.d/conda.sh
conda activate bfcl
cd env_service/launch_script
bash bfcl.sh

By default, the service will start on 127.0.0.1:8080.
If you need to change the host (--portal) or port (--port), edit the following line inside env_service/launch_script/bfcl.sh:

exec python -m env_service.env_service --env bfcl --portal 127.0.0.1 --port 8080

Feel free to replace the IP address or port number as needed to suit your environment.

Environment Service Interface

You can also launch the environment service manually without using a script by running:

python -m env_service.env_service [--args]

Here are the available command-line arguments:

Argument Type Default Description
--env str "appworld" The name of the environment to run.
--env_file_name str None Optional. Specific file name of the environment module. Defaults to "{env}_env.py".
--portal str "0.0.0.0" The IP address to bind the server to. Use "127.0.0.1" for local only.
--port int 8000 Port number to run the server on.
--debug bool False Whether to run the server in debug mode (True) or production mode (False), where useless logger will be removed in production mode.

Example

Start the server for appworld on port 8080 and localhost:

python -m env_service.env_service --env appworld --portal 127.0.0.1 --port 8080 --debug True

This gives you more flexibility to control and integrate the environment service in your own workflows.

API Interaction Flow

Our EnvService communicates via HTTP requests, primarily through the EnvClient located at env_service/env_client.py, which handles the service connection and has already been integrated into AgentEvolver.

  1. Query Task Candidates (Optional):
    POST /get_env_profile
    → Returns a list of available task_ids for a given environment type.

  2. Create Instance:
    POST /create
    → Initializes a new environment with env_type, task_id, and custom parameters.
    → Returns the initial state.

  3. Interact with Environment:
    POST /step
    → Sends an action to the environment instance.
    → Returns the next state, reward, and done signal.

  4. Evaluate Agent Behavior (Optional):
    POST /evaluate
    → Used for scoring overall task performance (e.g., episodic evaluation).

  5. Get Instance Information (Optional):
    POST /get_info
    → Retrieve metadata or progress information from the environment.

  6. Release Instance:
    POST /release
    → Frees the environment resources explicitly.

  7. Health Check:
    GET /healthz
    → Basic service status check.

Input & Output Data Structure

To help users interact effectively with the Appworld environment through the EnvService API, this section details the expected input and output data structures used in the /create and /step calls.

These examples illustrate what kind of requests to send and how to interpret the returned results. More examples can be found in env_service/interface.md

🛠️ /create – Environment Initialization

Description: Initializes a new environment instance and prepares its initial state.

🔽 Input Parameters

{
  "env": "appworld",
  "instance_id": "my-instance-id-001",
  "task_id": "task-001",
  "params": {
    "simple": false,
    "prompt": true
  }
}
Field Type Description
env string The environment to use. Set to "appworld".
instance_id string A user-defined identifier for tracking the environment instance.
task_id string The task assigned to this instance. Tasks can differ by difficulty or type.
params dict Additional configuration options:
simple bool – If true, uses a simplified interaction format.
prompt bool – If true, includes the environment prompt template in the state.

⏎ Output (Initial State)

{
  "state": [
    {
      "role": "system",
      "content": "[Structured prompt with task description and tool info]"
    },
    {
      "role": "user",
      "content": "[Task instruction from Appworld]"
    }
  ],
  "info": {
    "instance_id": "my-instance-id-001",
    "task_id": "task-001"
  }
}

🔄 /step – Sending an Action to the Environment

Description: Advances the environment one step using a user-generated action.

🔽 Input Format

{
  "action": {
    "role": "assistant",
    "content": "```python\nopen_app(\"Calendar\")\n```",
    "tool_calls": []
  }
}
Field Type Description
role str Typically "assistant", representing the agent performing the action.
content str The agent's natural language or code action. May include Python code blocks.
tool_calls list Optional list of structured tool calls (used for advanced interactions).

✅ You can omit tool_calls if your action is just a plain code string.

📤 Response from /step

{
  "state": [
    {
      "role": "assistant",
      "content": "Output:\n```\nApp opened successfully\n```"
    }
  ],
  "reward": 1.0,
  "is_terminated": true,
  "info": {}
}
Field Type Description
state list Environment’s response, typically including tool output or result summary.
reward float Task performance score. 1.0 = success, 0.0 = failure (configurable).
is_terminated bool Indicates whether the task is completed.
info dict Reserved for additional metadata (currently empty).

💡 Tips for Use

  • The content field should contain a well-formatted Python action or a command interpretable by Appworld.
  • You may use Markdown-style code blocks (```python) to wrap executable code for better readability.
  • If is_terminated: true, the environment expects no further actions.
  • Use the reward to evaluate success in training or evaluation settings.
  • The simple param in /create makes the prompt and output shorter — useful for debugging or fast testing.

For more details on tool calls, prompt generation, or evaluation metrics, refer to the Appworld Environment Internals section. """