AgentLab is a framework for developing and evaluating web agents on a variety of benchmarks supported by BrowserGym. This includes:

• WebArena
• WorkArena.L1, L2, L3
• VisualWebArena (coming soon...)
• MiniWoB

The framework enables the desing of rich hyperparameter spaces and the launch of parallel experiments using ablation studies or random searches. It also provides agent_xray, a visualization tool to inspect the results of the experiments using a custom gradio interface

Install agentlab

This repo is intended for testing and developing new agents, hence we clone and install using the -e flag.

git clone [email protected]:ServiceNow/AgentLab.git
pip install -e .

Set Environment Variables

export AGENTLAB_EXP_ROOT=<root directory of experiment results>  # defaults to $HOME/agentlab_results
export OPENAI_API_KEY=<your openai api key> # if openai models are used
export HUGGINGFACEHUB_API_TOKEN=<your huggingfacehub api token> # if huggingface models are used

Use an assistant to work for you (at your own cost and risk)

agentlab-assistant --start_url https://www.google.com

Prepare Benchmarks

Depending on which benchmark you use, there are some prerequisites

MiniWoB

export MINIWOB_URL="file://$HOME/dev/miniwob-plusplus/miniwob/html/miniwob/"

WebArena on AWS

TODO

WebArena on Azure

TODO

WorkArena

export SNOW_INSTANCE_URL="https://<your-instance-number>.service-now.com/"
export SNOW_INSTANCE_UNAME="admin"
export SNOW_INSTANCE_PWD=<your-instance-password>

Launch experiments

Create your agent or import an existing one:

from agentlab.agents.generic_agent.agent_configs import AGENT_4o

Run the agent on a benchmark:

study_name, exp_args_list = run_agents_on_benchmark(AGENT_4o, benchmark)
study_dir = make_study_dir(RESULTS_DIR, study_name)
run_experiments(n_jobs, exp_args_list, study_dir)

use agentlab.experiments.launch_commad.py to launch experiments with a variety of options. This is like a lazy CLI that is actually more convenient than a CLI. Just comment and uncomment the lines you need or modify at will (but don't push to the repo).

Debugging

For debugging, run experiments using n_jobs=1 and use VSCode debug mode. This will allow you to stop on breakpoints. To prevent the debugger from stopping on errors when running multiple experiments directly in VSCode, set enable_debug = False in ExpArgs

Parallel jobs

Running one agent on one task correspond to one job. When conducting ablation studies or random searches on hundreds of tasks with multiple seeds, this can lead to more than 10000 jobs. It is thus crucial to execute them in parallel. The agent usually wait on the LLM server to return the results or the web server to update the page. Hence, you can run 10-50 jobs in parallel on a single computer depending on how much RAM is available.

AgentXray

While your experiments are running, you can inspect the results using:

agentlab-xray

You will be able to select the recent experiments in the directory AGENTLAB_EXP_ROOT and visualize the results in a gradio interface.

In the following order, select:

• The experiment you want to visualize
• The agent if there is more than one
• The task
• And the seed

Once this is selected, you can see the trace of your agent on the given task. Click on the profiling image to select a step and observe the action taken by the agent.

Implement a new Agent

Get inspiration from the MostBasicAgent in agentlab/agents/most_basic_agent/most_basic_agent.py

Create a new directory in agentlab/agents/ with the name of your agent.

Misc

if you want to download HF models more quickly

pip install hf-transfer
pip install torch
export HF_HUB_ENABLE_HF_TRANSFER=1