AutoGPT Agent Server

This is an initial project for creating the next generation of agent execution, which is an AutoGPT agent server. The agent server will enable the creation of composite multi-agent systems that utilize AutoGPT agents and other non-agent components as its primitives.

Docs

You can access the docs for the AutoGPT Agent Server here.

Setup

We use the Poetry to manage the dependencies. To set up the project, follow these steps inside this directory:

0. Install Poetry

   pip install poetry

1. Configure Poetry to use .venv in your project directory

   poetry config virtualenvs.in-project true

2. Enter the poetry shell

   poetry shell

3. Install dependencies

   poetry install

4. Copy .env.example to .env

   cp .env.example .env

5. Generate the Prisma client

   poetry run prisma generate

   In case Prisma generates the client for the global Python installation instead of the virtual environment, the current mitigation is to just uninstall the global Prisma package:

   pip uninstall prisma

   Then run the generation again. The path should look something like this:
   <some path>/pypoetry/virtualenvs/backend-TQIRSwR6-py3.12/bin/prisma

6. Migrate the database. Be careful because this deletes current data in the database.

   docker compose up postgres redis -d
   poetry run prisma migrate dev 

Running The Server

Starting the server without Docker

Run the following command to build the dockerfiles:

poetry run app

Starting the server with Docker

Run the following command to build the dockerfiles:

docker compose build

Run the following command to run the app:

docker compose up

Run the following to automatically rebuild when code changes, in another terminal:

docker compose watch

Run the following command to shut down:

docker compose down

If you run into issues with dangling orphans, try:

docker compose down --volumes --remove-orphans && docker-compose up --force-recreate --renew-anon-volumes --remove-orphans  

Testing

To run the tests:

poetry run test

Development

Formatting & Linting

Auto formatter and linter are set up in the project. To run them:

Install:

poetry install --with dev

Format the code:

poetry run format

Lint the code:

poetry run lint

Project Outline

The current project has the following main modules:

blocks

This module stores all the Agent Blocks, which are reusable components to build a graph that represents the agent's behavior.

data

This module stores the logical model that is persisted in the database. It abstracts the database operations into functions that can be called by the service layer. Any code that interacts with Prisma objects or the database should reside in this module. The main models are:

• block: anything related to the block used in the graph
• execution: anything related to the execution graph execution
• graph: anything related to the graph, node, and its relations

execution

This module stores the business logic of executing the graph. It currently has the following main modules:

• manager: A service that consumes the queue of the graph execution and executes the graph. It contains both pieces of logic.
• scheduler: A service that triggers scheduled graph execution based on a cron expression. It pushes an execution request to the manager.

server

This module stores the logic for the server API. It contains all the logic used for the API that allows the client to create, execute, and monitor the graph and its execution. This API service interacts with other services like those defined in manager and scheduler.

utils

This module stores utility functions that are used across the project. Currently, it has two main modules:

• process: A module that contains the logic to spawn a new process.
• service: A module that serves as a parent class for all the services in the project.

Service Communication

Currently, there are only 3 active services:

• AgentServer (the API, defined in server.py)
• ExecutionManager (the executor, defined in manager.py)
• ExecutionScheduler (the scheduler, defined in scheduler.py)

The services run in independent Python processes and communicate through an IPC. A communication layer (service.py) is created to decouple the communication library from the implementation.

Currently, the IPC is done using Pyro5 and abstracted in a way that allows a function decorated with @expose to be called from a different process.

By default the daemons run on the following ports:

Execution Manager Daemon: 8002 Execution Scheduler Daemon: 8003 Rest Server Daemon: 8004

Adding a New Agent Block

To add a new agent block, you need to create a new class that inherits from Block and provides the following information:

• All the block code should live in the blocks (backend.blocks) module.
• input_schema: the schema of the input data, represented by a Pydantic object.
• output_schema: the schema of the output data, represented by a Pydantic object.
• run method: the main logic of the block.
• test_input & test_output: the sample input and output data for the block, which will be used to auto-test the block.
• You can mock the functions declared in the block using the test_mock field for your unit tests.
• Once you finish creating the block, you can test it by running pytest -s test/block/test_block.py.