The AI Agent Manifesto: The Path to Intelligent, Adaptable AI Agents

Introduction:

The evolution of artificial intelligence (AI) is advancing on two fronts: the pursuit of Artificial General Intelligence (AGI) and the development of autonomous agentic systems. While AGI aims to create large, multimodal language models (LLMs) that mimic human intelligence (I shall delve on this topic in a subsequent article), agentic systems focus on utilizing the reasoning and intelligence capabilities of? LLMs to achieve specific goals. In this article, we delve into the concept of "agents" in AI systems, exploring their characteristics and the path towards truly autonomous systems. More specifically, this article posits that agents encompass far more than just LLMs and their associated tools; they require a comprehensive set of characteristics, including memory, learning, planning, agency, and more.

The article discusses following key areas:

1. Agents and their characteristics
2. Anthropomorphizing AI Training for Memory
3. Memory and Learning Modules
4. Different Kinds of Memory systems
5. Planning Module
6. Configurability
7. Agency

I am happy to dedicate this article to Andrew Ng, whose pioneering work in artificial intelligence has inspired countless researchers and practitioners, including myself. His vision for democratizing AI education and his unwavering commitment to advancing the field have had a profound impact on the world.

What are Agents?

In the realm of AI, agents can be seen as the building blocks of autonomous systems — intelligent components capable of operating independently, similar to the fictional Skynet from the Terminator movies. Such a system is the ultimate goal of AI evolution. Unlike traditional computer systems that rely on human-defined inputs and outputs, agents possess a degree of autonomy. They are supposed to not only decide on the nature of inputs required but also determine the goals, adapting to environmental conditions and evolving workflows, as against conventional systems where inputs and outputs are predetermined.

Current implementations of agents, such as those found in frameworks like LangChain, often fall short of this ideal. While they typically leverage LLMs equipped with tools and planning modules like ReAct, they lack the true autonomy and adaptability that define fully realized agentic systems. To be considered a true agent, an entity must possess a broader set of capabilities beyond tool utilization and reasoning, including memory, learning, planning, and agency, as we will explore in the following sections.

Characteristics of AI Agents

AI agents must possess a range of characteristics:

1. Reasoning Module: A high-performing but light weight LLM with exceptional reasoning abilities shall be the central component. The LLM model should demonstrate a high MMLU (Massive Multitask Language Understanding) score of around 80%, indicating proficiency in various reasoning tasks. It need not store internet-scale data into its weights (refer my previous articles).?

2. Tools: Agents should have access to a diverse set of tools, APIs, and external databases. This includes the ability to code, test, and execute functions, allowing them to interact with the digital world effectively.? An agent would only be as good as the quality and breadth of the tools at its disposal, with the ideal scenario being a comprehensive library of tools and functions encapsulating every potential action within the agent's domain, leaving the LLM to focus only on reasoning and intelligent tool selection and execution towards achieving the goal.?

3. Planning Module: ?The planning module should go beyond the simple act-reason-observe cycle. It should be capable of planning entire workflows to achieve complex goals, determining the optimal sequence of actions, and dynamically adjusting the plans based on real-time feedback.

4. Memory Modules: Agents need a variety of memory modules, mirroring aspects of human memory. This includes a conversational buffer memory, short-term memory for recent events, long-term memory for knowledge and learning, entity memory for tracking objects and concepts, and contextual memory for understanding the current situation. More memory systems can be thought of depending on the domain and nature of goals involved.?

5. Storage and Processing: A significant portion of the agent's "brain" lies in its ability to intelligently process information from its memory modules. This involves filtering, prioritizing, and synthesizing information to make informed decisions and take appropriate actions.

6. Agency: True agency implies autonomous decision-making based on environmental cues and internal goals. Agents should be able to initiate actions without explicit human intervention, driven by their understanding of the situation and their objectives.

7. Learning:? Agents are not supposed to be designed for one-off tasks; they are meant to continuously learn and improve through repetitive execution of tasks within the autonomous system. This iterative learning process should refine their knowledge, skills, and decision-making capabilities.

8. Multimodal: Advanced agentic systems should be multimodal, capable of processing and understanding different types of data, including vision, speech, and text. This allows them to interact with the world in a more comprehensive and natural way. For example, a vision agent should be able to operate a CCTV camera rotating its position, zooming to specific objects in a way our human eyes explore the environment.?

9. Ethics and Safety Module: Agents should have an embedded module to ensure ethical decision-making and safety. This includes adhering to ethical guidelines, avoiding harmful actions, and recognizing potential biases in their reasoning and actions. This module should also ensure compliance with legal and regulatory standards.???

10. Sensory Integration: Where applicable, apart from multimodal capabilities, agents should be able to integrate sensory data seamlessly to form a coherent understanding of their environment. This includes processing inputs from various sensors (e.g., temperature, motion, and proximity sensors) to enhance situational awareness and decision-making.?

18. Resilience and Robustness:? Agents should be designed to be resilient against failures and robust in the face of unexpected situations. This includes having fail-safes, backup plans, and the ability to recover from errors gracefully.

19. Additional Characteristics:

???* Independence: Agents should operate independently, without requiring constant human supervision or intervention. This autonomy is crucial for their ability to handle tasks efficiently and effectively in dynamic environments. The concept of human input is not for verification or additional data, but should be used for feedback loops to further enhance learning opportunities

???* Adaptability: The ability to adapt to new or changing situations is a hallmark of intelligent agents. They should be able to learn from their experiences, adjust their strategies, and modify their behavior in response to unforeseen challenges.

???* Real-time and Batch Processing: Agents should be capable of both real-time processing for immediate responses and batch processing for tasks that require more time and computational resources. This flexibility allows them to handle a wide range of tasks efficiently.

???* Collaboration: In complex systems, agents often need to collaborate with each other to achieve shared goals. This requires effective communication, coordination, and the ability to negotiate and compromise when necessary.?

???* Persistence: Agents should be able to maintain their state and knowledge over time. This persistence ensures that they can resume tasks seamlessly after interruptions and build upon their previous experiences.?

???* Security:? Protecting the agent and the system from malicious attacks is crucial. Agents should be equipped with security measures to safeguard their data, prevent unauthorized access, and maintain the integrity of the system.?

The subsequent sections of this article elaborate more on the important topics of memory and learning, planning and agency aspects of agents.?

Anthropomorphizing AI Training for Memory - Training and the Emergence of Consciousness

Before delving into the nature and purpose of memory systems in AI agents, let's explore the intriguing concept of AI consciousness. While the debate continues on whether current leading GPTs truly possess consciousness, there's no denying that interacting with them feels akin to engaging with conscious entities. Regardless of the underlying reality, it's important to acknowledge the distinction between biological consciousness and the 'artificial consciousness' we seem to encounter in AI models like ChatGPT. The concept of anthropomorphizing AI, as explored in my previous LinkedIn posts, can help us navigate this complex landscape. While anthropomorphizing helps us understand and interact with AI entities exhibiting human-like traits, it can also shed light on their creation (as discussed in my four-part LinkedIn series on this topic). In this section, we'll extend this anthropomorphic lens to examine the process of training AI models.

The process of training a Large Language Model (LLM) involves optimizing its weights to achieve specific performance goals. Intriguingly, this process can be likened to the cultivation of a raw form of consciousness. When we train an LLM model by optimizing its weights, we are doing something similar to training the 'artificial consciousness' or 'digital brain' of the model.? Unlike traditional deep learning models with simpler structures, the multi-head attention mechanisms within Transformers, the building blocks of LLMs, create a complex network of learned representations. These representations, potentially involving multiple layers of interconnected weights, give rise to a multidimensional understanding of language and context.

During inference, the transformer's attention mechanism selectively engages different layers of weights based on the given prompts. This dynamic behavior suggests that the model continues to learn and adapt even with fixed weights, exhibiting a form of ongoing refinement.?

This phenomenon can be described as the "training of raw consciousness," where the model's inherent capacity for understanding and response evolves through interaction with data and the optimization process.

Consciousness and Intelligence: A Nuanced Relationship

When we talk about consciousness, we immediately think that it is synonymous with intelligence. However, consciousness is more about awareness and reasoning capability. A conscious object need not necessarily be a highly intelligent one.?

Within consciousness, there can be a spectrum ranging from minimally intelligent to highly intelligent states. This evolution of consciousness occurs through active learning and interaction with the environment. For instance, living organisms begin with minimal intelligence at birth, such as insects or fish. As they evolve over time, eventually becoming humans, their consciousness develops and becomes more intelligent. This intelligence is essentially the result of training consciousness over extended periods through numerous experiences with external systems. Another explanation is that consciousness needs to evolve before it can capture and process external stimuli in a meaningful way leading to higher degrees of intelligence. But higher states of consciousness require knowledge about external environment and experiences in dealing with it.?

Thus, while consciousness and intelligence are often conflated, they are not synonymous. Consciousness, the awareness of one's existence and surroundings, can exist with varying degrees of intelligence.? The evolution of consciousness is intrinsically linked to learning and adaptation, as organisms interact with their environment and acquire new knowledge and skills.?

A crucial component of consciousness that manifests intelligence is the memory system.

In the context of AI agents, intelligence can be viewed as a multifaceted construct encompassing several key abilities. One such ability is the power to reason and analyze information, enabling agents to make informed decisions and solve problems. Equally important is the capacity for memory, which acts not only as a repository of past experiences but also as a fundamental mechanism for learning from those experiences. Memory systems enable AI agents to store and recall information, allowing them to adapt and improve their performance over time. By integrating advanced memory systems into AI agents, we can create a framework for them to expand their consciousness, mirroring the way humans learn and grow through accumulated knowledge and experiences. This integration allows AI agents to develop a deeper understanding of their environment, enhance their decision-making capabilities, and ultimately, exhibit more intelligent behavior.

Consciousness and Intelligence: A Nuanced Relationship, Illustrated by TinyLlama

An interesting example of the distinction between consciousness and intelligence can be found in the AI model experiment, TinyLlama. This 1-billion parameter model, which shares the architecture of LLaMa 2, was designed to test the limits of reasoning capabilities for the smallest possible model. While TinyLlama could answer some advanced questions well, it struggles with common-sense reasoning (refer to my earlier LinkedIn post on this topic).

At first glance, the answers TinyLlama provides might suggest a lack of emergent abilities or digital consciousness due to its size. However, a closer examination reveals a different story. TinyLlama's responses demonstrate a form of reasoning, though not always aligned with real-world facts. This indicates a form of consciousness with a lesser degree of intelligence. The model's overtraining, which prevented the creation of additional layers for reasoning due to limited parameters, and possibly corrupted existing weights meant for reasoning, could explain this phenomenon. This situation can be likened to a degradation of intelligence or consciousness, where the ability to reason effectively is compromised.

TinyLlama serves as a compelling example that consciousness and intelligence are not the same. This distinction underscores the importance of not only developing conscious AI agents but also ensuring they possess the necessary knowledge and reasoning capabilities to function intelligently in the real world.

An important part of extension of consciousness by learning can be achieved through memory systems.?

Expanding Consciousness: Beyond Fine-Tuning and Prompt Engineering

If we want to create LLM based 'Agents' that can learn over a period of time, it means, in an anthropomorphic sense, we want to train the 'consciousness component’ of that LLM further. In other words, these agents can be equated to 'trained consciousness components' that evolve over a period of time based on learning from several tasks executed in the past. (Note: the RLHF training of models like ChatGPT is not just about alignment, it plays a significant part in this dimension).??

While techniques like fine-tuning or prompt engineering can refine an LLM's performance on specific tasks, they are insufficient for truly expanding the 'raw artificial consciousness' and fostering a higher degree of general intelligence. These methods essentially tweak the existing knowledge and capabilities embedded within the model's fixed weights, but they don't fundamentally alter its underlying cognitive architecture or capacity for broader understanding.

To achieve a more profound evolution of artificial consciousness, we need to look beyond these methods. The key lies in augmenting the model's capabilities in a way that mirrors the development of human intelligence. Just as our brains grow and learn through experience, accumulating knowledge and skills over time, AI agents need a mechanism to expand their awareness and understanding beyond the initial training data.

The solution, as we'll explore further, lies in developing sophisticated memory and learning systems that enable agents to continuously adapt and evolve, building upon their existing knowledge base and refining their decision-making processes. This approach takes inspiration from the way the human brain processes and integrates new information, gradually expanding its cognitive repertoire.?

The Path Forward: Evolving Agent Intelligence through Memory and Learning

While the initial training of an LLM lays the foundation for its "raw consciousness," the journey towards true artificial intelligence doesn't end there. To create agents that can adapt, learn, and evolve, we must go beyond simply retraining the model for every new situation. Instead, we need to focus on developing sophisticated memory and learning mechanisms that empower the agent to expand its knowledge and capabilities independently.

It's important to distinguish between two distinct facets of AI development:

1. Training for Emergent Consciousness: This involves the initial training process where the LLM's underlying architecture and parameters are established. It's during this phase that the potential for "raw consciousness" emerges, shaping the model's fundamental understanding and reasoning abilities.
2. Training for Memory and Learning: This ongoing process focuses on enhancing the agent's intelligence by equipping it with the tools to learn from experience, store and retrieve information, and adapt its decision-making processes. This involves developing sophisticated memory modules, feedback loops, and learning algorithms that enable the agent to continuously evolve and refine its knowledge base.

By recognizing this distinction, we can approach the development of AI agents more strategically. Rather than solely focusing on refining the initial "consciousness" through retraining, we can invest in building powerful memory and learning systems that empower agents to become truly intelligent and adaptable. This approach not only enhances the agent's performance and capabilities but also aligns with the way biological organisms learn and evolve, gradually expanding their cognitive repertoire through experience and interaction with the world.

Ultimately, by fostering this ongoing evolution of agent intelligence through memory and learning, we can unlock the full potential of AI, creating autonomous agents that can navigate complex challenges, adapt to changing circumstances, and contribute meaningfully to solving real-world problems.

Training for Memory and Learning??

Though we widely use the term 'Machine Learning,' the reality, as of today, is that we haven't truly invented a system that learns continuously over time (though it's unclear what kind of memory & learning system is already in place with models like ChatGPT). The intelligence of the current GPT models are fixed with the weights of the models. Any further learning achieved through prompt engineering like in-context learning are not persistent and they are not available at a later point of time. If the AI system has to learn over time based on execution of tasks over time, it needs to store its learning as part of the system. This storage of learnings can be termed as memory systems of agents.?

An agent's memory should extend beyond mere conversational history or transaction records. Instead, it should be viewed as an evolving component of the model's reasoning capabilities. While real-time RLHF training for memory and learning might be impractical, the goal is to enhance the original LLM's consciousness through an extension focused on these aspects.

While ideally, this memory extension would be trained in tandem with the original LLM, a more practical approach is to conceptualize it as a separate component that generates additional prompt instructions based on the LLM's learned experiences. Rather than continuously retraining the entire model, we focus on enhancing its memory, allowing the existing 'consciousness' to work with an ever-growing knowledge base.

Each task the agent performs is analyzed for learning opportunities. When identified, these learnings are transformed into prompt instructions, embedded, and stored in a persistent database. When the agent encounters a similar task in the future, relevant learnings are retrieved and integrated into the existing prompt, providing contextual guidance. The agent itself determines which past experiences are most pertinent, constructing a dynamic context that enhances its decision-making and performance over time.

Processing Memory

Over a period of time, the learning could become wieldy and voluminous. At the initial stage, the agent might store every possible learnings into the persistent storage and fetch them to its? buffer memory based on relevance. Over a period of time, some of the learnings might never be used at all. And some of the learnings might be precious and required for some critical actions but that are carried out very rarely. So, the agent needs to decide on weeding out unnecessary memory on its own so that its buffer memory consisting of learned context is not filled with unnecessary learnings. Hence, it should be able to process the memory so that it can decide to remove some memory altogether, store some in a short term memory module and keep some in a long term memory module which may be kept outside the system in cheaper storage. This memory processing should be happening as a background process.

A mechanism of feedback loop is essential for such memory based systems. This feedback could be provided as a batch process which might require additional system or manual intervention or as part of the completion of the task itself which depends on the nature of the tasks.?

Thus, the learnings of an Agentic system can be compared to the way humans learn. Every interaction with the external environment is stored as memory in our brain. Some of these are immediate buffer memory, some are short term memory while a large part is long term memory. The memory flows starting from buffer state to long term memory over a period of time. Not all memory (all interactions with the external environment) reaches the long term storage in our brain. During the REM sleep, our brain decides which of the memory is worth to be stored in short term memory, which part of buffer should be thrown out then and there and which part of short term memory should be moved to long term memory. In the similar way, the agent system can process its memory through some background batch processes.?

To conclude, instead of training the model for extending its 'digital brain or consciousness', we aim to focus on its memory which becomes a component to work on by the existing consciousness. So, every task performed by the agent since its inception gets analyzed for learning opportunities. If a learning opportunity is found, it should be converted into prompt instructions converted into an embedding that is stored in a persistent database which should be fetched every time the same task or similar task is performed by that agent.?

This additional context, based on memory and learning, which evolves over time, is presented alongside the existing prompt instructions to perform the specific task. The agent should be able to decide what part of the previous learnings are relevant for the current task and construct the context accordingly.????

Different Kinds of Agent Memory Systems??

For autonomous agents to function effectively, a robust and diverse set of memory systems is essential. These systems enable agents to process information, learn from experiences, and interact intelligently with their environment. Below are the different kinds of memory systems crucial for autonomous agents:

1. Short-Term Memory

Short-term memory allows autonomous agents to temporarily store and manage information needed for immediate tasks and quick decision-making. This type of memory is essential for processing information in real-time, such as intermediate steps in problem-solving, executing a series of commands, or handling transient data during computations. It acts as a workspace for the agent to manipulate and use information actively before either discarding it or transferring it to long-term memory.

2. Long-Term Memory

Long-term memory is critical for retaining information over extended periods. This memory system allows agents to store vast amounts of knowledge, including facts, learned skills, and experiences. Long-term memory enables agents to draw on past interactions and accumulated knowledge to improve performance and make informed decisions over time.

3. Conversational Memory

Conversational memory is a specialized type of memory that allows agents to remember details from previous interactions with users. This includes remembering user preferences, past questions, and relevant context from prior conversations. Conversational memory is vital for creating a personalized and seamless user experience, as it helps agents maintain continuity and build rapport with users.

4. Entity Memory

Entity memory focuses on storing and recalling information about specific entities, such as people, places, and objects. This type of memory helps agents keep track of relevant details associated with different entities, enabling more accurate and context-aware interactions. For example, in a customer service scenario, entity memory allows an agent to remember customer details, purchase history, and preferences.

5. Contextual Memory

Contextual memory stores information specific to the current context or situation in which the agent is operating. This includes environmental variables, ongoing tasks, recent interactions, and other relevant factors that influence decision-making and behavior. Contextual memory enables agents to adapt their responses and actions based on real-time situational awareness, enhancing their effectiveness and responsiveness in dynamic environments.

Including contextual memory ensures that agents can dynamically adjust their behavior and decision-making processes based on the changing circumstances they encounter. This capability is crucial for real-world applications where environmental conditions and user interactions vary continuously.

6. Procedural Memory

Procedural memory enables agents to remember how to perform specific tasks and actions. This type of memory is crucial for executing complex procedures and routines that have been learned over time. Procedural memory supports the automation of repetitive tasks and ensures that agents can carry out actions consistently and efficiently.

7. Episodic Memory

Episodic memory allows agents to recall specific events and experiences from the past. This memory system helps agents understand the context of past interactions and use this information to inform future actions. Episodic memory is important for learning from experiences, adapting to new situations, and improving decision-making processes.

8. Semantic Memory

Semantic memory stores general knowledge about the world, including concepts, facts, and meanings. This type of memory allows agents to understand and use language effectively, make inferences, and apply knowledge to various situations. Semantic memory is essential for enabling agents to process information accurately and respond appropriately to diverse queries.

Integration of Memory Systems

For autonomous agents to operate effectively, these different memory systems must be seamlessly integrated. The interplay between short-term and long-term memory ensures that immediate needs are met while building a knowledge base for future use. Conversational and entity memory systems enhance personalized interactions, while procedural, episodic, and semantic memories contribute to the agent's overall intelligence and adaptability.

By incorporating these diverse memory systems, autonomous agents can achieve a higher level of functionality, learning, and intelligence, ultimately leading to more effective and human-like interactions.

Planning: The Architect of Agentic Workflows

The planning module in an AI agent is akin to the architect of its cognitive processes and workflow. However, it should go far beyond the current implementations found in frameworks like LangChain, which primarily focus on planning individual tasks or goals. To achieve true autonomy, the planning module should be capable of orchestrating entire workflows, akin to a conductor leading an orchestra of specialized agents.

Beyond Task-Level Planning

Current planning modules often operate within a limited scope, guiding the agent through a single task's act-reason-observe cycle. While this is valuable for specific actions, it falls short of enabling the agent to tackle complex, multi-faceted objectives. Imagine an agent tasked with planning a marketing campaign. A traditional planning module might focus on a single aspect, like crafting an email. However, a more sophisticated module would envision the entire campaign, from market research and audience segmentation to content creation, distribution, and performance analysis.?

To achieve this level of planning, the agentic system itself should act as the conductor, orchestrating the actions of various specialized agents. Each agent, like a musician in an orchestra, plays a specific role in achieving the overall goal. The planning module determines the composition of the orchestra, assigning tasks to different agents based on their expertise, and coordinating their efforts to ensure a harmonious outcome.?

The planning module should not be limited to executing pre-defined workflows. Instead, it should be capable of dynamically designing workflows based on the specific goals and constraints of the situation. For example, tt should be able to decide whether a hierarchical, parallel, or sequential approach is most appropriate, and whether a manager LLM is necessary to oversee and coordinate the activities of other agents.

The Limitations of Current Approaches: Predefined Workflows

Current implementations of planning modules, even in advanced frameworks like CrewAI, often rely on developers to explicitly define the structure of workflows. This involves making predetermined decisions about the necessity of a manager LLM, the sequencing of tasks (sequential, parallel, or hierarchical), and the specific interactions between agents. This approach restricts the agent's adaptability and autonomy, as it is confined to operating within the boundaries of these predetermined guidelines.

In an ideal scenario, the agentic system itself should be capable of making these decisions based on the specific context and goals. For instance, if the task is relatively simple and self-contained, the agent might determine that a linear workflow is sufficient. However, for more complex tasks with multiple interdependent components, a hierarchical or parallel approach might be more effective.

Similarly, the decision of whether to employ a manager LLM should not be pre-determined. In some cases, a manager might be essential to oversee the coordination of multiple agents, while in others, a decentralized approach might be more efficient. By empowering the agentic system to make these decisions dynamically, we can create agents that are truly adaptable and capable of handling a wider range of tasks and scenarios.

The Need for Higher-Level Planning

This higher-level planning capability is crucial for tackling complex, real-world problems. For instance, an agent tasked with managing a customer support system could autonomously handle a wide range of responsibilities. This might include answering customer inquiries through various channels (email, chat, phone), resolving technical issues, processing returns and refunds, escalating complex cases to human agents, and analyzing customer feedback to identify areas for improvement. Each of these responsibilities could be delegated to specialized agents with expertise in specific areas. The planning module would then orchestrate the activities of these agents, ensuring that customer inquiries are handled promptly and effectively, resources are allocated efficiently, and the overall customer experience is optimized. Depending on the nature of the problem, the agent might decide to adopt a different type of workflow.

By evolving beyond task-level planning and embracing the role of a workflow orchestrator, the planning module can unlock the true potential of agentic systems. This will enable AI agents to tackle complex challenges, adapt to dynamic environments, and collaborate seamlessly to achieve ambitious goals. As AI technology continues to advance, we can expect planning modules to become increasingly sophisticated, ultimately paving the way for a new generation of autonomous systems that can truly transform our world.

Configurability of AI Agents

A fundamental characteristic of effective AI agents is their high degree of configurability, enabling customization of the parameters that govern their behavior, learning, and memory processes. However, current agent frameworks like CrewAI and Autogen, while offering pre-built components or black-box architectures, often lack the flexibility required to adapt to the diverse demands of real-world scenarios.

The Limitations of Fixed and Black-Box Approaches

Fixed agentic components, though convenient, may not align with the unique requirements of different tasks and environments of the real world. They lack the flexibility to evolve and optimize their performance based on changing circumstances. Similarly, black-box architectures obscure the inner workings of the agent's memory and learning mechanisms, making it difficult to understand, debug, or tailor them to specific needs.

The Need for Tailored Memory Systems

The concept of memory itself can vary significantly across different applications. Some tasks might require a strong emphasis on short-term memory for immediate responses, while others might prioritize long-term memory for storing accumulated knowledge. Additionally, the nature of entities tracked, the type of contextual information relevant, and the retention periods for different memory types can all vary depending on the specific domain and use case.

Empowering Adaptability Through Configurability

To address these challenges, AI agents should be designed with a high degree of configurability, especially concerning their memory systems. This includes:

* Memory Module Sizes: The ability to adjust the capacity of different memory modules (conversational buffer, short-term, long-term, entity, contextual) allows for fine-tuning based on the task's memory requirements.

* Retention Periods:? Customizable retention periods for each memory type enable the agent to balance the need for retaining relevant information with the need to manage memory resources efficiently.

* Learning Rates and Strategies:? Different learning tasks might necessitate varying learning rates and strategies. Configurable learning parameters empower the agent to adapt its learning process to the specific problem domain.

* Feedback Mechanisms:? The effectiveness of feedback can depend on the nature of the task and the environment. Configurable feedback mechanisms allow for tailoring the feedback loop to optimize the agent's learning and performance.

By offering such a high level of configurability, developers and users can mold AI agents into powerful tools that cater to their specific needs. This not only enhances the agent's performance and efficiency but also fosters a deeper understanding of its inner workings, promoting transparency and trust in its decision-making processes.

Agency: The Driving Force of Autonomous Action

The concept of agency is pivotal in understanding the behavior and capabilities of AI agents. While consciousness encompasses various dimensions, the fundamental aspect of agency lies in the agent's ability to interact with its environment and make autonomous decisions.

Levels of Consciousness and the Role of Agency?

Consciousness encompasses various dimensions and levels. At its core, consciousness involves awareness of the external environment (or input stimuli in the case of GPT-based AI models). How an agent processes these stimuli and interacts with its surroundings is determined by its intelligence, memory, and, in the case of biological organisms, emotions. In its simplest form, consciousness is merely this awareness.

The second, more complex level of consciousness involves awareness of one's own existence as a distinct entity, a concept often referred to as self-awareness. While basic consciousness exists without self-reflection, the second level introduces the ability to introspect and recognize oneself as separate from the environment.

My intuition is, ignoring the concept of 'Embodied AI' (refer to my earlier post on LinkedIn on this topic), artificial consciousness may not reach the second level of consciousness where it gains awareness of its own existence. But more importantly, we simply don't require agents to have such self-awareness. The basic primary purpose of such awareness in a biological organism is to ensure its own survival. Once it cognizes the external environment and about itself, its emotions become the torch bearer for survival with regard to actions it should take. These emotions determine which actions to take and in what circumstances. Ignoring AGI for a moment, we dont require Agentic systems to have emotions - that level of subjectivity is not required for business applications. The intuition is while biological consciousness might always come with emotions, this may not be the case with artificial consciousness. Similarly, the consciousness with 'self awareness' is also not required for agentic systems. However, a certain amount of 'agency', arising out of the first level of consciousness should be necessary and sufficient for an agentic system

To conclude, the primary purpose of self-awareness in biological organisms is often linked to survival and the pursuit of well-being. This drives emotions, which in turn influence decision-making. However, for agentic systems in business applications, emotions and subjective biases can be detrimental. Similarly, the self-preservation instincts associated with self-awareness might not be relevant for agents designed for specific tasks.

The Agency of Cells: A Model for AI Agents

A more appropriate model for agency in AI systems can be found in the cells of our bodies. These cells, while lacking self-awareness, exhibit a form of agency. They interact with their environment, taking in nutrients, producing energy, and carrying out specific functions according to their genetic programming. They have a defined scope of action and operate within specific boundaries. This form of agency, although lacking subjectivity, enables them to fulfill their roles effectively.

Defining Agency in AI Agents

Artificial consciousness in agentic systems can be conceptualized at a similar level to cellular consciousness, but with significantly enhanced learning and memory capabilities. An AI agent should be able to perceive its environment, determine relevant inputs, and set its own goals based on its understanding and objectives. Unlike human agency, which is often driven by survival and happiness, agent agency is task-oriented and focused on achieving specific outcomes.

The Current Gap in Agency

Current definitions and implementations of AI agents often fall short of incorporating this crucial aspect of agency - the current agents are nothing more than LLMs with some tools - there is really no agency involved. Many agents are simply programmed to follow pre-defined instructions or react to specific stimuli. While they might exhibit some degree of autonomy in their actions, they lack the ability to set their own goals and determine the best course of action to achieve them.

The Path Forward: Empowering Agents with True Agency

To unlock the full potential of agentic systems, we must empower them with true agency. This involves developing agents that can:

Perceive and Interpret: Accurately perceive and interpret information from their environment, including sensor data, user inputs, and feedback.

Set Goals: Autonomously define goals based on their understanding of the situation, task requirements, and available resources.

Plan and Execute: Devise effective plans and strategies to achieve their goals, selecting appropriate actions and adapting to changing circumstances.

Learn and Adapt: Continuously learn from their experiences, refining their knowledge, skills, and decision-making capabilities.

Collaborate: Interact and collaborate with other agents to achieve shared objectives, leveraging their collective intelligence and capabilities.

By incorporating these elements of agency, we can create AI agents that are not merely tools but truly autonomous entities capable of making independent decisions and taking proactive actions to achieve their goals.

Some Additional Considerations?

* Configurability and Modularity: While modularity is desirable for flexibility, the emphasis should be on designing agents that are highly configurable rather than relying on pre-built libraries of agents (I believe Autogen leans towards black-box, pre-built agents). Configurable agents,? as against prebuilt black boxes, allows for greater adaptability and customization to specific tasks and environments. If an agent needs updating or modification, its internal components can be reconfigured, eliminating the need for complete replacement. This approach also promotes transparency and avoids the limitations of black-box solutions.

* Goal Specificity: Agents and agentic systems should be built with specific goals in mind rather than being generic, one-size-fits-all solutions. This goal-oriented approach ensures that agents are equipped with the necessary knowledge, skills, and tools to excel in their designated domains. For example, an agent in an medical agentic system designed for medical diagnosis would have different capabilities and training data compared to an agent in an financial agentic framework designed for financial analysis. The Agentic Frameworks should provide higher level abstractions that allows for easy building of agents with configurable parameters.?

* Avoid Black-Box Approaches:? Transparency and explainability are paramount in the development of AI agents. Black-box approaches, where the internal workings of the agent are opaque and difficult to understand, should be avoided. This is because black-box models make it difficult to diagnose errors, identify biases, and ensure that the agent's decisions align with human values and ethical principles.

Conclusion

The development of truly autonomous AI systems hinges on the creation of intelligent, adaptable agents. By understanding and implementing the characteristics discussed in this article, we can pave the way for a future where AI agents collaborate seamlessly to achieve complex goals, ultimately transforming industries and enhancing our lives.

While the path to fully autonomous agents is still ongoing, the advancements made in AI research and development bring us closer to this reality. As we continue to explore the possibilities of agentic systems, it is crucial to prioritize transparency, ethical considerations, and the development of agents that augment human capabilities rather than replace them.