#multiagent

[timofeyzinin]

so i accidentally deployed 6 AI agents to production at 2am and now they argue with each other about code formatting. every. single. time. a PR gets opened. one wants 4 spaces. another wants tabs. the third says "just use prettier and stop asking questions." the fourth is silent but keeps approving things. they've been running for 72 hours now. yesterday they started a debate about whether comments lie. i think they're right. most comments do lie. anyway, everything's fine. that's just how you learn about agent orchestration. by letting them orchestrate you.

[newstech24]

How startups can obtain outsized outcomes by leveraging multi-agent techniques

In March, AWS introduced the final availability of its new multi-agent capabilities, bringing the expertise into the fingers of companies throughout virtually each trade. Till now, organizations have largely relied on single-agent AI techniques, which deal with particular person duties however typically wrestle with complicated workflows. These techniques can even break down when companies…

How startups can obtain outsized outcomes by leveraging multi-agent techniques

[mysocial8onetech]

MetaGPT: A Framework for Multi-Agent Meta Programming

Find out how MetaGPT compares to other frameworks such as AutoGPT and AgentVerse in terms of capabilities and performance.

MetaGPT is a framework that enables you to use natural language to create and execute meta programs for multi-agent collaboration and coordination. Meta programs are programs that can generate or modify other programs based on some input or context. In this article, you will find out how MetaGPT works, what are its key features and capabilities, and how it compares to other frameworks.

[wrongbat]

Agent class using PyZqm and tornado ioloop

I created a simple class that will be the parent class for all the various types of agents in my multi-agent simulation. The class inherits from Process, so it needs to define a run() method. This run method will be called when the process is started. 

Agent has an abstract method, setup(), the implementation of which is done in the sub-class. In setup() the sockets are created, and the agent sends initial requests/init messages.

I’ll figure out how to deal with heart-beating stuff and reliability later. For now this class will do for prototyping applications.

Here’s the code:

import zmq from zmq.eventloop import ioloop, zmqstream from multiprocessing import Process import abc from datetime import datetime class Agent(Process): __metaclass__ = abc.ABCMeta def __init__(self, name): Process.__init__(self) self.context = zmq.Context() self.loop = ioloop.IOLoop.instance() def run(self): self.say('Setting up agent...') self.setup() self.say('Starting ioloop...') self.loop.start() @abc.abstractmethod def setup(self): return def say(self, msg): print('{} - {}: {}'.format(datetime.now().strftime('%H:%M:%S'), self.name, msg)) class Server(Agent): def handle_socket(self, msg): address, m = msg[0], msg[2] self.say(m) self.socket.send_multipart([address, '', 'REQUEST OK']) def setup(self): self.socket = self.context.socket(zmq.ROUTER) self.socket.bind('tcp://*:6000') stream_pull = zmqstream.ZMQStream(self.socket) stream_pull.on_recv(self.handle_socket) class Client(Agent): def handle_socket(self, msg): self.say(msg) self.socket.send_multipart(["", 'NEW REQUEST']) def setup(self): self.socket = self.context.socket(zmq.DEALER) self.socket.connect('tcp://localhost:6000') stream_pull = zmqstream.ZMQStream(self.socket) stream_pull.on_recv(self.handle_socket) self.socket.send_multipart(["", 'NEW REQUEST']) if __name__ == '__main__': Server(name = 'server').start() Client(name = 'client').start()

[wrongbat]

Writing a wrapper function WITHOUT using decorators

When I first started learning about Python decorators I found them pretty confusing an unintuitive, and it took me a while before I started using them. What I often want to do is modify a function with some extra functionality without having to add the extra code in my original function. Decorators are excellent for doing that, but if you find them confusing (which I did) then you can write a wrapper like this:

def wrap(agent): assert hasattr(agent, 'loop') lfunc = agent.loop def before_loop(): print('Before loop') def after_loop(): print('Doing stuff after loop') def wrapped(*args, **kwargs): before_loop() try: lfunc(*args, **kwargs) except Exception, e: print(e) after_loop() agent.loop = wrapped class TestAgent: def loop(self, n): for i in range(n): print('In loop') print(x) if __name__ == '__main__': a = TestAgent() wrap(a) a.loop(5)

if you run the script you will see that the before_loop and after_loop are also executed:

Before loop In loop In loop In loop In loop In loop Doing stuff after loop

The try/except block in wrapper() makes sure that after_loop() is called even if loop() throws an exception.

You should not forget to use a copy of the loop function (third line in the example above) as it will cause a RuntimeError due to infinite recursion. So this will crash:

def wrap_broken(agent): assert hasattr(agent, 'loop') def before_loop(): print('Before loop') def after_loop(): print('Doing stuff after loop') def wrapped(*args, **kwargs): before_loop() agent.loop(*args, **kwargs) after_loop() agent.loop = wrapped class TestAgent: def loop(self, n): for i in range(n): print('In loop') if __name__ == '__main__': a = TestAgent() wrap_broken(a) a.loop(5)

[aa-nplus1]

AgentWare Workshop 

//AA Visiting school, Rovinj, Croatia //Alisa Andrasek ++ Jose Sanchez //6 days of generative design/code