Earlier this year, we saw many autonomous AI agents — the likes of BabyAGI, AutoGPT, MetaGPT and AgentGPT — mushrooming in the space and performing different tasks. The latest one to join the party is Microsoft’s AutoGen.

AutoGen is far more advanced than the above mentioned examples as it can use multiple AI agents to solve a task. This is in contrast to most other AI agents, which are single agent frameworks and can only use a single AI agent to solve a task. 

AutoGen facilitates interactions between multiple agents, enabling them to collaborate and complete tasks. These AutoGen agents are adaptable, conversational, and operate in diverse modes, incorporating LLMs, human inputs, and various tools as needed.

But why do we really care about autonomous AI agents?

Passive AI vs Autonomous AI 

Today, the most common problem one faces with LLMs is that they hallucinate and lack reasoning capabilities. Be it ChatGPT or any other LLM model, they currently need to be trained by humans through methods like fine-tuning or reinforcement learning. They don’t self evolve. To put it simply, they need human intervention. 

On the other hand, Autonomous AI agents can work discreetly within LLMs to enhance their reasoning abilities. To simplify, imagine several distinct AI agents engaging in conversations, exchanging advice before an LLM formulates a response. They are able to perceive their surroundings, make decisions, and take actions on their own. 

For instance, self-driving cars are autonomous AI agents that can navigate roads and avoid obstacles without human input. They use a variety of sensors, such as cameras, radar, and lidar to perceive their surroundings and make decisions.

In contrast, passive AI lacks the ability to learn on the fly. Passive AI systems operate based on pre-programmed rules or static datasets and cannot adjust their behavior in response to changing circumstances. 

“This fundamental difference in adaptability and responsiveness sets autonomous agents apart from passive AI systems, making them better suited for tasks requiring real-time decision-making in complex and evolving environments,” said Anand Trivedi, head of artificial intelligence at Aavenir. 

ChatGPT users may have noticed that ChatGPT shows steps while solving math problems, even when it’s not necessary. With Autonomous AI agents, users won’t need to see what ChatGPT is doing behind the scenes. Instead, they can directly view the result while ChatGPT interacts with the agent. Enhancing math capabilities is just one application of AutoGen. 

With the help of multi-agent conversations, it is possible for chat-optimised LLMs like GPT-4 to take feedback, allowing LLM agents to work together in conversations. For example, agents can exchange reasoning, observations, critiques, and validation in a dialogue involving humans or other agents. On similar lines, Meta released Shepherd, a language model explicitly tuned to critique model-generated outputs.

When appropriately configured with the right prompts and inference settings, a single LLM can demonstrate a wide array of capabilities. Conversations among agents with different configurations can effectively amalgamate these diverse LLM abilities in a modular and complementary manner.

Microsoft’s AutoGen uses LLMs to create smart agents that can do various tasks. These agents can play roles, understand context, learn from conversations, receive feedback, adapt, and even code. These abilities can be combined in new ways to make the agents more skilled and independent. 

Moreover, AutoGen also has features to improve LLMs, like caching results and handling errors. In some cases, human involvement is important. AutoGen allows humans to join agent conversations and provide input when needed, depending on the agent’s settings.

Future possibilities

It is widely speculated that in the future, AI agents, if not replacing human jobs entirely, will certainly become new companions in the workplace. With their ability to think and make decisions autonomously, the need for human presence in the workplace is expected to be significantly reduced. This will enable businesses to operate more efficiently while remaining competitive.

Unlike humans, autonomous AI agents do not require sleep or lunch breaks. They can work 24/7, ensuring effective production, faster results, and reducing the tedious task workloads of current employees. 

Moreover, as autonomous AI agents become deeply familiar with organisational processes, it gains the ability to quickly detect process violations and suggest improvements in workflows, compliance procedures, and more. With autonomous AI agents, self-evolving models can surpass the generally available models in the market. Over time, they might even begin to develop strategies for the organisations. 

Interestingly, rumors abound about OpenAI possibly revealing their first entirely autonomous agent at the upcoming OpenAI DevDay conference. This speculation is closely tied to OpenAI’s recent acquisition of Global Illumination, where AI agents are undergoing training within a gamified simulation.