agency defines a common communication and action framework for creating AI agent integrated applications.

The library provides a low-level means for connecting agents, systems, and human users by defining actions, callbacks, and access policies that you can use to monitor, control, and interact with your agents.

agency handles the details of the common messaging system and allows discovering and invoking actions across parties, automatically handling things such as reporting exceptions, enforcing access restrictions, and more.

• Straightforward class/method based agent and action definition
• Supports defining single process applications or networked agent systems using AMQP

• Before/After action callbacks for observability or other needs
• Access policies and permission callbacks for access control

• Multithreaded (though python's GIL is a bottleneck for single process apps)
• AMQP support for multiprocess and networked systems (avoids GIL)
• Python multiprocess support is planned for better scalability on single-host systems

• Not yet developed, but is planned

• Two OpenAI agent examples
• HuggingFace transformers agent example
• Simple Flask/React web interface included
• Direct host access for agents
• Docker configuration for reference and development

agency is an implementation of the Actor model for building AI agent integrated systems.

In agency, all entities are represented as instances of the Agent class. This includes all humans, software, and AI-driven agents that may communicate as part of your application.

All agents may expose "actions" that other agents can discover and invoke at run time. An example of a simple agent implemention could be:

class CalculatorAgent(Agent):
  def _action__add(a, b):
    return a + b

This defines an agent with a single action: "add". Other agents will be able to call this method by sending a message to an instance of CalculatorAgent and specifying the "add" action.

other_agent._send({
  'to': 'CalcAgent',
  'thoughts': 'Optionally explain here',
  'action': 'add',
  'args': {
    'a': 1,
    'b': 2,
  },
})

Actions must also specify an access policy, allowing you to control access for safety. For example:

  @access_policy(ACCESS_PERMITTED) # This allows the action at any time
  def _action__add(a, b):
    ...

You can also define before/after and permission callbacks for various purposes:

class CalculatorAgent(Agent):
  ...
  def _before_action(self, original_message: dict):
    # Called before any action is attempted

  def _after_action(self, original_message: dict, return_value: str, error: str):
    # Called after any action is attempted

  def _request_permission(self, proposed_message: dict) -> bool:
    # Called before an ACCESS_REQUESTED action is attempted for run-time review

A Space is how you connect your agents together. An agent cannot communicate with others until it is added to a common "space".

There are two included Space implementations to choose from:

• NativeSpace - which connects agents within the same python process
• AMQPSpace - which connects agents across processes and systems using an AMQP server like RabbitMQ.

Here is an example of creating a NativeSpace and adding two agents to it.

space = NativeSpace()
space.add(CalculatorAgent("CalcAgent"))
space.add(AIAgent("AIAgent"))
# The agents above can now communicate

These are just some of the main agency features. For more detailed information please see the docs directory.

pip install agency

or

poetry add agency

To run the demo, please follow the directions at examples/demo. After a short boot time you can visit the web app at http://localhost:8080 and you should see a simple chat interface.

The following is a screenshot of the web UI that demonstrates the multiple demo agents intelligently interacting and following orders.

There are two OpenAI based agents: "FunctionAI" and "CompletionAI", named for the API's they use, and "Chatty" a simple chat agent who uses a small local transformers based model for demonstration.

The screenshot also demonstrates the results of rejecting an action and directing an agent to use a different approach in real time. After I explained my rejection of the read_file action (which happened behind the scenes on the terminal), "FunctionAI" appropriately used the shell_command action with wc -l Dockerfile.

Though you could entirely create a simple agent using only the primitives in agency (see examples/demo/agents/), it is not intended to be a full-fledged agent toolset.

Projects like LangChain and others are exploring how to create purpose-built agents that solve diverse problems using tools. agency is focused on the problems surrounding agent/tool/human integration, such as observability and access control.

More likely, you would use LangChain and other libraries for defining agent behavior and rely on agency to provide the connective layer for bringing agents and other systems together.

So in comparison, agency is a smaller but more general purpose application framework compared to libraries like LangChain that focus on enabling individual agent behavior.

• It's a new project, so keep that in mind in terms of completeness, but see the issues page for what is currently planned, and the Roadmap below for the high level plan.

• This library makes use of threads for each individual agent. Multithreading is limited by python's GIL, meaning that if you run a local model or other heavy computation in the same process as other agents, they may have to wait for their "turn". Note that I/O does not block, so networked backends or services will execute in parallel.

  For blocking processes, it's recommended to use the AMQPSpace class and run heavy computations in isolation to avoid blocking other agents.

• This API does not assume or enforce predefined roles like "user", "system", "assistant", etc. This is an intentional decision and is not likely to change.

  agency is intended to allow potentially large numbers of agents, systems, and people to come together. A small predefined set of roles gets in the way of representing many things generally. This is a core feature of agency: that all entities are treated the same and may be interacted with through common means.

  The lack of roles may require extra translation code when integrating with role based APIs. See the implementation of OpenAIFunctionAgent for an example.

• There is currently not much by way of storage support. That is mostly left up to you and I'd suggest looking at the many technologies that focus on that. The Agent class implements a simple _message_log array which you can make use of or overwrite to back it with longer term storage. More direct support for storage APIs will likely be considered in the future.

Please do!

git clone [email protected]:operand/agency.git
cd agency
poetry install

See the demo directory for instructions on how to run the demo.

The demo application is written to showcase both native and AMQP spaces and several agent examples. It can also be used for experimentation and development.

The application is configured to read the agency library source when running, allowing changes to be tested manually.

Ensure you have Docker installed. A small RabbitMQ container will be automatically created.

You can run the test suite with:

poetry run pytest

• Multiprocess Support: An additional space type utilizing python multiprocessing, as another parallelism option for single-host systems.

• Multimodal Support: Image/audio transfer for use with multimodal models or other multimedia services.

• More Examples: More examples of integrations with popular AI libraries and tools such as Langchain and oobabooga.

Please see the issues page.

If you have any suggestions or otherwise, feel free to add an issue!