by Adi Prerepa and Andrew Han
AutoInfraGPT was an attempt to create an LLM agent (using an AGI architecture that we built ahead of time) with the capability of creating an EC2 instance on Amazon, SSHing into it, and then running a curl command on the EC2 instance to a separate server.
The main logic of the code is shown below:
actions = [
Action(
name="create_ec2_instance",
description="Create an EC2 instance on Amazon. Returns the public ipv4 of this instance.",
function=create_ec2_instance,
),
Action(
name="ssh_and_execute_cmd",
description="SSH into an instance and run a command.",
function=ssh_and_execute_cmd,
),
]
objective = f"""Create 1 EC2 Instance with name rat. Ssh into that instance and make HTTP request via curl to http://54.245.39.114:8080. Print the response from the server. Declare done."""
agent = AgenticGPT(
objective, actions_available=actions, model="gpt-4"
)
agent.run()where create_ec2_instance is a function we wrote to spin up an AWS VM and ssh_and_execute_cmd is another function we wrote to SSH into the created instance and run a UNIX command.
The agent is instantiated with actions and given the objective to "Create 1 EC2 Instance with name rat. Ssh into that instance and make HTTP request via curl to http://54.245.39.114:8080. Print the response from the server. Declare done."
The hardcoded endpoint hosts a Python server which returns a simple HTML. GPT-4 and AgenticGPT should, given the instructions, know to call the two functions in order, resulting in a curl reuquest to the given endpoint. See the output below to see the output from the agent.
This was a trivial example that involved two steps, with calls to two functions. However, you could imagine exposing many more of AWS's APIs to AgenticGPT, giving it the ability to use more resources, read logs, and respond accordingly. Taken as a whole, we could conceivably see a world in which devops are automated by agents.
-
git clone https://github.com/handrew/agentic_gpt -
Move
autoinfra_gpt.pyinto the cloned folder (sorry!AgenticGPTis not yet on pypi) -
Set up a
.envfile withOPENAI_API_KEYwith your API key andSSH_KEYFILEwith AWS SSH key. -
cdinto the folder and thenpython autoinfra_gpt.py. You have to have AWS credentials on your local machine.
INFO:agentic_gpt.agent.agentic_gpt:Taken steps 1 of maximum 100.
INFO:agentic_gpt.agent.agentic_gpt:
Thoughts: I need to create an EC2 instance first. I will use the create_ec2_instance action with the instance_name as 'rat' and instance_count as 1.
INFO:agentic_gpt.agent.agentic_gpt:Reasoning: The first step in the objective is to create an EC2 instance with the name 'rat'. This can be achieved using the create_ec2_instance action.
INFO:agentic_gpt.agent.agentic_gpt:Chosen action: create_ec2_instance
INFO:botocore.credentials:Found credentials in shared credentials file: ~/.aws/credentials
INFO:botocore.credentials:Found credentials in shared credentials file: ~/.aws/credentials
INFO:agentic_gpt.agent.agentic_gpt:Completed action create_ec2_instance. Result: ec2-54-213-51-236.us-west-2.compute.amazonaws.com
INFO:agentic_gpt.agent.agentic_gpt:Taken steps 2 of maximum 100.
INFO:agentic_gpt.agent.agentic_gpt:
Thoughts: Now that the EC2 instance is created, I need to SSH into it and make an HTTP request to the provided URL. I will use the curl command for this.
INFO:agentic_gpt.agent.agentic_gpt:Reasoning: The next step in the task is to SSH into the created EC2 instance and make an HTTP request. The curl command is a tool to transfer data from or to a server, using on
e of the supported protocols. In this case, we are using it to make an HTTP request.
INFO:agentic_gpt.agent.agentic_gpt:Chosen action: ssh_and_execute_cmd
sshing with: {}, username {} ec2-54-213-51-236.us-west-2.compute.amazonaws.com ec2-user
INFO:paramiko.transport:Connected (version 2.0, client OpenSSH_8.7)
INFO:paramiko.transport:Authentication (publickey) successful!
INFO:agentic_gpt.agent.agentic_gpt:Completed action ssh_and_execute_cmd. Result: <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN" "http://www.w3.org/TR/html4/strict.dtd">
<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<title>Directory listing for /</title>
</head>
<body>
<h1>Directory listing for /</h1>
<hr>
<ul>
<li><a href=".bash_history">.bash_history</a></li>
<li><a href=".bash_logout">.bash_logout</a></li>
<li><a href=".bashrc">.bashrc</a></li>
<li><a href=".cache/">.cache/</a></li>
<li><a href=".profile">.profile</a></li>
<li><a href=".python_history">.python_history</a></li>
<li><a href=".ssh/">.ssh/</a></li>
<li><a href=".sudo_as_admin_successful">.sudo_as_admin_successful</a></li>
</ul>
<hr>
</body>
</html>
INFO:agentic_gpt.agent.agentic_gpt:Taken steps 3 of maximum 100.
INFO:agentic_gpt.agent.agentic_gpt:
Thoughts: I have created the EC2 instance and made the HTTP request via curl. The response from the server is stored in the memory. Now, I need to print the response and declare the task as done.
INFO:agentic_gpt.agent.agentic_gpt:Reasoning: The task requires me to print the response from the server after making the HTTP request. The response is already stored in the memory as a result of the ssh_and_execute_cmd action. After printing the response, I will declare the task as done.