Quick Start

Previous Next

This guide provides two distinct paths for training an agent:

  1. Basic GRPO Training: A standard method to get started quickly.
  2. AgentEvolver Training: An advanced method that supports a self-evolving agent training.

Prerequisites: One-Time Global Setup

Before you begin, run these commands in your terminal to configure your environment. You only need to do this once.

  1. Initialize Conda
source <YOUR_CONDA_PATH>/etc/profile.d/conda.sh
  1. Configure API Endpoints
export DASHSCOPE_API_KEY="<YOUR_API_KEY>"
export HF_ENDPOINT=https://hf-mirror.com

💡 Tip: Add the export commands to your ~/.bashrc or ~/.zshrc file to set them automatically in new terminal sessions.

Part A: Basic GRPO Training

Step 1: Setup Env-Service (AppWorld for example)

This launches the simulation environment (e.g., AppWorld) where the agent will operate. This service will run in the background. You'll need a new terminal for the next step.

conda activate appworld
bash env_service/launch_script/appworld.sh

Step 2: Start Basic GRPO Training

This command starts the training process using the GRPO method.

conda activate agentevolver
bash examples/run_basic.sh

Part B: AgentEvolver Training

Step 1: Setup Env-Service (AppWorld for example)

Just like in basic training, this launches the agent's simulation environment. This service will run in the background. You'll need a new terminal for the next step.

conda activate appworld
bash env_service/launch_script/appworld.sh

Step 2: Setup ReMe-Service

This service gives the agent long-term memory and the ability to reflect on past actions. This service will listen for requests on http://127.0.0.1:8001. Keep this terminal open.

Configure API Endpoints:

export FLOW_EMBEDDING_API_KEY="<YOUR_API_KEY>"
export FLOW_EMBEDDING_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
export FLOW_LLM_API_KEY="<YOUR_API_KEY>"
export FLOW_LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1

conda activate reme
cd external/reme
reme \
  config=default \
  backend=http \
  thread_pool_max_workers=256 \
  http.host="127.0.0.1" \
  http.port=8001 \
  http.limit_concurrency=256 \
  llm.default.model_name=qwen-max-2025-01-25 \
  embedding_model.default.model_name=text-embedding-v4 \
  vector_store.default.backend=local \
  op.rerank_memory_op.params.enable_llm_rerank=false

Step 3: Start AgentEvolver Training

With the environment and ReMe services running, start the AgentEvolver training.

conda activate agentevolver
bash examples/run_overall.sh

Part C: Multi-Node Training

Please read Part A and Part B first.

To run AgentEvolver and train agents on multiple nodes, we need some extra setup.

Step 0. Prepare the Training

Follow the instructions in Configuration to prepare the training environment on all nodes.

Step 1. Start the Ray cluster

In Part A and Part B, we leave the Ray cluster to be started by AgentEvolver. To enable multi-node training, we need to start the Ray cluster manually.

Start Ray on all nodes:

# all ray node must be in the same conda environment
conda activate agentevolver

# node 1 as head
ray start --head

# other nodes as followers
ray start --address='<head addr>'

Step 2. Start Env-Service (AppWorld for example)

Start envservice on one node:

conda activate appworld
cd env_service
bash launch_script/appworld.sh

Once you see the successful startup logs, you're good to go. The address of this node will be used in the next step.

If ReMe-Service is needed, start it on one node.

Step 3. Start AgentEvolver

Create a launch script based on run_basic.sh or run_overall.sh:

  • Update env_url to the address of the envservice node
  • Adjust other training parameters as needed

Then start training from one node:

conda activate agentevolver
bash run_basic.sh

AgentEvolver will automatically use the Ray cluster and spread the training tasks to all nodes.

Configuration Environment Service