Part 1: The Foundations of Autonomy: Defining the Mind and Body

The fields of artificial intelligence and robotics are built on a foundational duality: the separation of the “mind” (the decision-making intelligence) from the “body” (the system that acts in the world). To achieve true autonomy, a system must have both: a “body” to interact with its environment, which is the Autonomous System, and a “mind” to direct its actions, which is the Intelligent Agent. This report will first define these core components before exploring the advanced research that is fusing them together.

1.1 What is an Autonomous System? Clearing the Confusion

The term “Autonomous System” (AS) is a source of significant confusion, as it holds distinct meanings in at least three separate, high-level fields.¹

In Mathematics: An autonomous system is a system of ordinary differential equations, often expressed as dx/dt = f(x), in which the laws governing the system’s evolution do not explicitly depend on the independent variable, such as time.¹ This is distinguished from a non-autonomous system, dx/dt = f(x, t), where time is a direct factor.² In essence, a mathematical AS’s future state depends only on its current state.

In Internet Networking: An autonomous system refers to a collection of IP networks and routers under the control of a single administrative entity.¹ This is a core concept for global internet routing (Border Gateway Protocol), where each AS (e.g., a large ISP or tech company) manages its own routing policies.

In AI and Robotics: An autonomous system refers to an autonomous robot or entity that can perform desired tasks in unstructured environments “without continuous human guidance”.¹

While these definitions appear unrelated, the unifying theme is “independence” or “self-governance.” The mathematical system is independent of time, the internet system is administratively independent from other networks, and the robotic system is operationally independent from a human controller. For the remainder of this report, “Autonomous System” will refer exclusively to the AI and robotics definition: a physical or software system, such as a robot, that is designed to act on its own.⁴

1.2 The Mind of the Machine: The Intelligent Agent (IA)

The “mind” that drives an autonomous system is known as an Intelligent Agent (IA). In artificial intelligence, an IA is defined as any entity that perceives its environment, takes actions autonomously to achieve specific goals, and may improve its performance by learning.⁵

This concept can be simplified:

• An Agent is anything that “perceives” its environment using sensors (e.g., a robot’s cameras, a software program’s data feed) and “acts” upon that environment using actuators (e.g., a robot’s wheels, a program’s recommendations).⁵
• An Intelligent Agent is one that does so in a way that is purposeful and goal-directed.

This concept is so fundamental that the entire field of Artificial Intelligence is often defined as “the study and creation of these rational agents”.⁵ Intelligent agents exist on a vast spectrum of complexity, from a simple thermostat (which senses temperature and acts to maintain a goal) to a self-driving car (which senses a complex 3D environment and acts to navigate safely to a destination).⁶

1.3 The Agent’s Motive: The Objective Function

An agent’s behavior is not random; it is guided by an objective function (also called a goal function), which “encapsulates their goals”.⁵ This function serves as a “measure of success,” and a Rational Agent is defined as one that strives to achieve the best possible outcome as defined by this measure.⁵

This objective function can have different “flavors” depending on the field of study, but the concept remains the same:

• Utility Function: In economics, this represents an agent’s desirability or happiness with a particular state.⁵
• Reward Function: In reinforcement learning, this provides feedback (rewards or punishments) that the agent learns to maximize over time.⁵
• Loss Function: In machine learning, this is a measure of error that the agent seeks to minimize.⁵
• Fitness Function: In evolutionary systems, this determines an agent’s success at survival and reproduction.⁵

It is critical to separate an agent’s “Why” (its goal) from its “How” (its methodology). The objective for a vacuum robot is its “Why”: “clean the entire floor.” The methodology it uses is its “How.” This “How” can be a simple reflex (bouncing off walls) or a highly complex, model-based approach (building a digital map of the room). The goal is the same, but the intelligence of the method is different.

1.4 Table 1: The Spectrum of Agent Intelligence (The How)

The “How” of an agent is determined by its internal architecture. These architectures represent a ladder of increasing intelligence, allowing agents to move from simple reactions to complex, goal-oriented reasoning.⁵

Source: Synthesized from classifications in ⁵

As we move up this spectrum, we arrive at agents that are not just reactive, but deliberative. This forms the basis of the most advanced research in agent-based AI.

Part 2: Agentic AI: The Architecture of a Deliberative Mind

The term “Agentic AI” represents the pinnacle of this spectrum. It is not a fundamentally different concept from an Intelligent Agent, but rather a specialized class of IA defined by its high degree of autonomy and its focus on complex, multi-step problem-solving.

2.1 The Next Leap: Defining Agentic AI

Agentic AI refers to a class of artificial intelligence that focuses on autonomous systems capable of pursuing “complex goals” with “limited or no human intervention”.⁶ The key capability of an agentic system is that it does not merely perform a single, pre-defined task. Instead, it can “independently analyze challenges, develop strategies and execute” complex, multi-step plans to achieve a high-level goal.¹⁰ This requires “sophisticated reasoning and iterative planning”¹⁰ that mimics “human decision-making”.¹¹ To accomplish this, an agentic system must integrate and orchestrate various AI techniques, such as natural language processing (NLP), machine learning (ML), and computer vision.⁸

2.2 The Critical Distinction: Agentic AI vs. Robotic Process Automation (RPA)

The most effective way to understand what Agentic AI is is to contrast it with what it is not. Its foil is Robotic Process Automation (RPA), a common business technology.

• RPA: Automates “rule-based, repetitive tasks with fixed logic”.⁸ An RPA bot is a rigid script.
• Agentic AI: “Adapts based on data inputs” and changing conditions.⁸ An agentic bot is a flexible problem-solver.

This represents a fundamental shift from automating a procedure to automating an outcome. Consider the task of paying an invoice. An RPA bot is given a procedure: “IF you receive an invoice, THEN open it, extract the number from line 10, and paste it into the payment system.” If a new invoice format arrives where the total is on line 11, the RPA bot will fail. Its “fixed logic”⁸ has broken. An Agentic AI is given a goal: “Pay all invoices.” It receives the new invoice. Using NLP and ML⁸, it analyzes the document, reasons that the word “Total” is next to the number on line 11, and adapts its strategy. It formulates a new plan to complete its goal, even in the face of an unexpected change. This is the core of agentic behavior.

2.3 A Model for the Machine Mind: Beliefs, Desires, and Intentions (BDI)

To achieve this level of flexible reasoning, an agent needs a “mental” framework. The Belief-Desire-Intention (BDI) model is one of the “best known and best studied models” for this purpose.¹² It was directly inspired by a philosophical model of human practical reasoning, developed by Michael Bratman.¹³ The core purpose of the BDI software model is to allow an agent to “balance the time spent on deliberating about plans (choosing what to do) and executing those plans (doing it)”.¹⁴ Crucially, it provides a mechanism for “separating the activity of selecting a plan… from the execution” of that plan.¹⁴ This separation is what prevents the agent from being paralyzed by indecision, allowing it to commit to a course of action.

2.4 Table 2: Deconstructing the Agent’s Mind (BDI Model)

The BDI model breaks down an agent’s “thought process” into distinct, interacting components.¹⁴

Source: Synthesized from ¹⁴

2.5 Why Intention is the Key to Autonomy

The “Intention” component is the magic ingredient that elevates an agent from being merely reactive (like a thermostat⁶) to being truly agentic. It provides stability and focus. A simple agent’s “Desire” (e.g., “set temperature to 70”) is fixed and directly triggers an action. It does not deliberate. A BDI agent, however, deliberates.

• It holds multiple, often conflicting, Desires (e.g., Desire 1: “Save energy,” Desire 2: “Keep human comfortable”).
• It consults its Beliefs to get context (e.g., Belief 1: “It is 3:00 AM,” Belief 2: “No human is in the room,” Belief 3: “Electricity is cheap”).
• Based on this deliberation, it commits to an Intention¹⁵: “I will prioritize ‘Save energy’ and let the temperature drop slightly.”

This “Intention” is now a stable, committed goal.¹⁴ The agent does not re-evaluate all its desires every second. Instead, it focuses on finding and executing the Plan to achieve its intention. This ability to deliberate, commit, and then focus on execution is what allows an agent to pursue the complex, long-term goals⁸ that define agentic AI.

Part 3: Autonomous Systems in Practice: From Single Robots to Swarms

This section connects the abstract “mind” (the Agentic AI) to the physical “body” (the Autonomous System or Robot). It explores how these agents are “embodied” in the physical world and how they interact with each other to form “swarms.”

3.1 Embodied Agents: The Autonomous Robot

An autonomous robot is the physical manifestation of an intelligent agent. The concept that formally links them is the “Embodied Agent”.⁵ An embodied agent is an intelligent agent that “interacts with the environment through a physical body within that environment”.⁵ Mobile robots are a prime example of a physically embodied agent.⁴ In this relationship:

• The Intelligent Agent (the “mind”) is the core software that perceives, reasons, and decides.⁵
• The Autonomous Robot (the “body”) is the collection of sensors (cameras, lidar), actuators (wheels, arms), and hardware that allows the agent to execute its decisions in the physical world.⁴

3.2 The Four Pillars of Physical Autonomy

For a physical robot to be truly autonomous, its agent “mind” must manage four distinct capabilities.⁴

• Self-Maintenance (Proprioception): The ability to sense its internal state.⁴ “Proprioception” is the agent sensing itself. The classic example is a robot sensing its own low battery and autonomously navigating to a charging station.⁴
• Environmental Sensing (Exteroception): The ability to sense the external world.⁴ “Exteroception” is the agent sensing its surroundings to perform tasks and “stay out of trouble”.⁴
• Task Performance: The ability to perform physical tasks. This ranges from simple (a vacuum robot cleaning⁴) to complex conditional tasks. A conditional task requires the agent to respond differently based on conditions, such as a security robot⁴ that “respond[s] in a particular way depending upon where the intruder is.”
• Autonomous Navigation: The ability to know where it is (localization) and move “point-to-point” without human guidance.⁴ This is used in indoor navigation (hospital robots⁴) and complex outdoor navigation (Mars rovers⁴).

An embodied agent’s “Beliefs” (from the BDI model) are thus built from two distinct channels: the outside world (exteroception) and its own body (proprioception). True autonomy requires the agent to reason over both simultaneously. It must be able to decide, “My Intention is to go to the kitchen, but my Belief is that my battery is at 10% (proprioception) and there is a wall in the way (exteroception). Therefore, I must generate a new plan: go to the charger first, avoiding the wall.”

3.3 A Showcase of Autonomous Robots (Real-World Examples)

These embodied agents are already deployed across numerous sectors⁴:

• Historical: The earliest examples of long-range autonomous robots are space probes.⁴
• Domestic: The most common examples are self-driving vacuums.⁴ More advanced systems include Amazon’s Astro, designed for home monitoring and eldercare.⁴
• Industrial: Robot arms on assembly lines are considered autonomous, though their autonomy is restricted by their “highly structured environment” and inability to move.⁴
• Transportation: This includes self-driving cars⁴ and SpaceX’s “Autonomous Spaceport Drone Ships,” which are ocean-going platforms that position themselves to safely land and recover Falcon 9 rockets at sea.⁴
• Scientific: The Mars rovers (MER-A Spirit and MER-B Opportunity) autonomously “compute optimal paths” on the fly by mapping the 3D surface and calculating safe routes to their destinations.⁴
• Military: The South Korean SGR-A1 is an autonomous sentry gun developed to assist troops in the Korean Demilitarized Zone. It is a “highly classified” system that integrates “surveillance, tracking, firing, and voice recognition”.⁴

3.4 The Collective: Multi-Agent Systems (MAS)

Autonomy is not limited to a single agent. A Multi-Agent System (MAS) is a computerized system composed of “multiple interacting intelligent agents”.¹⁶ The core purpose of a MAS is to solve problems that are “too difficult or impossible for a single agent or a monolithic system” to handle.¹⁶ The agents within a MAS can be software programs¹⁶, robots¹⁶, or even “combined human-agent teams”.¹⁶

These systems are defined by three key characteristics¹⁶:

• Autonomy: Agents are at least partially independent.
• Local Views: No single agent has a “full global view” of the entire system.
• Decentralization: There is no “boss” agent designated as the central controller.

The power of a MAS comes from the concept of emergence. A MAS can “manifest self-organisation” and “complex behaviors even when the individual strategies of all their agents are simple”.¹⁶ The classic example is a flock of birds¹⁶: no single bird is “smart” or knows the flock’s shape. Each bird follows simple rules (e.g., “stay close to neighbors, don’t crash”). But from these simple micro-level rules, the complex, intelligent, and fluid macro-level behavior of the flock “emerges.” The “intelligence” of the system is not in any single agent but in their interactions.

3.5 Table 3: MAS (Engineering) vs. ABM (Science)

The research makes a subtle but critical distinction between two types of multi-agent studies, which is vital for understanding their purpose.¹⁶

Source: Synthesized from ¹⁶

In short, MAS is prescriptive: it’s about building a system to do something. ABM is descriptive: it’s about building a simulation to understand something. The perfect real-world example synthesizes all three concepts: Waymo’s “Carcraft” simulation environment.¹⁶

• Carcraft is an Agent-Based Model (ABM). It is a simulation used for scientific insight, modeling the behavior of human drivers and pedestrians.¹⁷
• Waymo uses this ABM to build and test its Intelligent Agent (IA) (the “mind” of a single self-driving car).
• The ultimate goal is to deploy a Multi-Agent System (MAS) (a network of autonomous cars that coordinate on the road), which is an engineering solution to transportation.¹⁶

Part 4: The Path Forward: Real-World Challenges and Critical Concerns

While the capabilities of agentic and autonomous systems are advancing rapidly, the research also highlights profound real-world challenges, risks, and ethical dilemmas that define the frontier of the field.

4.1 Why Autonomy Is So Hard: The Brittleness Problem

The primary barrier to widespread, safe autonomy is brittleness. Autonomous robots are “highly vulnerable to unexpected changes in real-world environments”.⁴ This is not just a problem of handling large, obvious obstacles. The research states that “Even minor variations like a sudden beam of sunlight disrupting vision systems… can cause entire systems to fail”.⁴ This brittleness exists because robotics is an “inherently systems problem,” where a failure in any single module (perception, planning, or actuation) can “compromise the whole robot”.⁴ This vulnerability is a symptom of the “open-world” challenge. Agents are often trained on “datasets captured under controlled conditions” and “struggle to generalize” to the messy, dynamic real world.⁴ They fail when encountering “unknown objects, occlusions,” and “rapidly changing environments”.⁴

The “beam of sunlight” example⁴ reveals that the biggest challenge is not just the agent’s “mind” (planning) but its “senses” (perception). The perception module, when blinded by the sun, delivers a false Belief to the agent (e.g., “The road ahead is clear”). The agent’s planning module might be perfect, but it is now acting on corrupted information, leading to a catastrophic failure. This “reality gap”⁴ between simulation and the real world means the research frontier is focused not just on making smarter agents, but more robust agents that can engage in “self-supervised, lifelong learning”⁴ to adapt to the open world.

4.2 Critical Concerns: Security, Privacy, and Ethics

As agents become more autonomous, they introduce new and scalable risks.

Security and Privacy:

Research into new “agentic web” products (like Microsoft’s NLWeb⁸) has already raised specific alarms.⁸

• Data Exfiltration: These products have been “criticised for exfiltrating information about their users to third-party servers”.⁸
• Security Flaws: They “expos[e] security issues” because “the way the agents communicate often occur through non-standard protocols”.⁸

This risk is scalable precisely because of the agent’s autonomy. A human using a web browser is predictable. An autonomous agent, given a goal like “Find me the cheapest flight,” might autonomously share the user’s entire profile (their “Beliefs” and “Desires”) with dozens of servers simultaneously to achieve its objective, using its own “non-standard protocols” that bypass traditional firewalls.

Ethical Dilemmas:

The most profound ethical questions concern “Lethal Autonomous Weapons Systems” (LAWS).⁴ This is not a theoretical, futuristic concern. It is a present-day reality, embodied by systems like the SGR-A1 autonomous sentry gun.⁴ This “highly classified”⁴ system, with its integrated “surveillance, tracking, firing, and voice recognition” capabilities, moves the “human-in-the-loop” to “human-on-the-loop,” or potentially, out of the loop entirely. These developments are the subject of ongoing discussions at the United Nations, highlighting the “societal and economic impacts”⁴ as autonomy becomes more pervasive.

4.3 The Future Outlook: A Fusion of Mind and Language

The future of autonomous systems is being defined by the fusion of classic agent architectures with the power of modern Large Language Models (LLMs). Research into “language model-based multi-agent systems” has been identified as a “new area of research” and a “new paradigm” for application development.¹⁶

This fusion represents the report’s culminating point: the future of Agentic AI appears to be an LLM-based BDI model.

• The classic BDI model¹⁴ provides the time-tested architecture (Beliefs, Desires, Intentions).
• The LLM provides the powerful reasoning engine.

In this new paradigm:

• Beliefs: The LLM’s vast, pre-trained knowledge acts as the agent’s foundational “Belief” set.
• Desires: The user’s prompt (e.g., “Plan a 5-day trip to Paris, book flights, and find a hotel”) becomes the agent’s “Desire.”
• Deliberation & Planning: The LLM’s “sophisticated reasoning and iterative planning”¹⁰ is the deliberation process, weighing options and formulating a strategy.
• Intention: The final, coherent, step-by-step plan that the LLM generates—and commits to executing—is the “Intention.”

This fusion of the stable, logical BDI framework with the generative and flexible reasoning power of LLMs is what bridges the gap from simple logic to the “human-like decision-making”¹¹ that defines the agentic revolution. As research solves for “fault-tolerance”¹⁶ and “multi-robot coordination”⁴, this new model points toward a future of increasingly capable and integrated autonomous systems.