Fact-checked by Grok 2 weeks ago

Software agent

A software agent is a self-contained computer program designed to perceive its environment, make decisions, and perform actions autonomously to achieve predefined goals on behalf of a user or another entity.^[1] These agents operate persistently without requiring continuous human intervention, distinguishing them from traditional software by their ability to delegate high-level tasks and adapt to dynamic conditions.^[1] The concept of software agents emerged from the field of distributed artificial intelligence in the late 1970s and gained prominence in the 1990s, building on foundational models like Carl Hewitt's Actor formalism, which emphasized concurrent, autonomous computational entities.^[2] Key characteristics defining agenthood include autonomy, where agents control their own actions and internal state without direct human oversight; social ability, enabling interaction with other agents and humans to collaborate on tasks; reactivity (or responsiveness), allowing timely perception and response to environmental changes; and pro-activeness, which drives goal-directed behavior and initiative-taking beyond mere reactions.^[3] These properties, formalized in influential works from the mid-1990s, provide a framework for designing agents that exhibit intelligent, flexible behavior in complex systems.^[3] Software agents vary in sophistication and application, ranging from simple reactive agents that respond to stimuli without internal planning to deliberative agents employing symbolic reasoning for long-term goal pursuit.^[2] Common types include interface agents for user assistance, mobile agents that migrate across networks, and collaborative agents that coordinate in multi-agent systems to solve distributed problems.^[2] They have been applied in domains such as personal information management, electronic commerce, computer games, and process automation, where their autonomy enhances efficiency and scalability.^[1] Ongoing research continues to integrate learning capabilities, enabling agents to improve performance over time through adaptation.^[2]

Definition and Fundamentals

Core Definition

A software agent is defined as an autonomous entity that perceives its environment through sensors and acts upon that environment through actuators, selecting actions to maximize its expected performance measure in pursuit of designated goals.^[4] This framework, central to artificial intelligence, views agents as programs or systems capable of rational behavior by processing perceptual inputs to generate appropriate outputs.^[4] At its core, a software agent comprises key components: perception mechanisms to gather environmental data, decision-making processes to evaluate options against goals, action execution via interfaces with the environment, and often learning capabilities to improve performance over time through adaptation or experience.^[4] These elements enable the agent to function independently within its operational context, whether in simulated or real-world settings.^[4] The scope of software agents encompasses both reactive agents, which respond directly to environmental stimuli, and deliberative agents, which engage in planning and reasoning to anticipate future states.^[5] Key defining properties include autonomy, allowing operation without constant human intervention; proactivity, enabling goal-directed initiative; and social ability, facilitating interaction with other agents or users.^[5] Reactivity, the capacity to perceive and timely respond to environmental changes, further supports these traits.^[5] Software agents range from simple implementations, such as rule-based controllers mimicking basic feedback loops like a coded thermostat, to more advanced intelligent variants that incorporate sophisticated reasoning and learning.^[4] While simple agents suffice for straightforward tasks, intelligent agents exhibit greater flexibility and adaptability, distinguishing them in complex, dynamic environments.^[5]

Key Characteristics

Software agents are distinguished by several core properties that enable them to function effectively in dynamic environments. Central to their design is autonomy, the capacity to make decisions and take actions independently without requiring continuous human oversight, often guided by predefined goals to achieve specific objectives. This property allows agents to control their internal states and behaviors, distinguishing them from passive scripts or tools that depend on explicit user directives.^[5] Complementing autonomy are reactivity and pro-activeness, which together provide flexibility in responding to and influencing the environment. Reactivity enables agents to perceive changes in their surroundings through sensors or data inputs and respond in real-time, such as a monitoring agent detecting system anomalies and initiating alerts or corrections. Pro-activeness, on the other hand, empowers agents to anticipate future states and initiate actions proactively to pursue goals, rather than merely reacting to stimuli—for instance, an e-commerce agent might predict inventory shortages and reorder supplies ahead of demand spikes. These traits ensure agents exhibit goal-directed behavior in partially observable or uncertain settings.^[5]^[2] Social ability further enhances agent functionality by facilitating interactions with other agents, humans, or systems in multi-agent environments, often through standardized communication protocols like agent-communication languages. This property supports collaboration, negotiation, and coordination, as seen in distributed systems where agents exchange information to optimize collective outcomes, such as in supply chain management. Meanwhile, adaptability and learning allow agents to evolve over time by incorporating machine learning techniques, refining their decision-making based on experience and feedback; learning agents, for example, adjust strategies through reinforcement learning to improve performance in evolving scenarios.^[5]^[6] Finally, rationality underpins agent behavior, defined as selecting actions that maximize expected utility given the agent's percepts, knowledge, and goals, thereby achieving optimal or near-optimal results. In practice, this often manifests as bounded rationality, where agents operate under computational constraints and incomplete information in complex real-world environments, balancing efficiency with effectiveness rather than pursuing perfect solutions. These characteristics collectively enable software agents to address tasks requiring persistence, intelligence, and interaction beyond traditional programming paradigms.^[6]^[7]

Historical Development

Early Concepts and Origins

The early concepts of software agents emerged from pre-1950s theoretical foundations in cybernetics and automata theory, which provided the intellectual groundwork for autonomous computational entities. Norbert Wiener's 1948 book Cybernetics: Or Control and Communication in the Animal and the Machine defined cybernetics as the study of control and communication in machines and living organisms, emphasizing feedback loops to enable self-regulation and adaptation—key principles for systems that could perceive, act, and adjust independently.^[8] Complementing this, Alan Turing's 1936 paper "On Computable Numbers, with an Application to the Entscheidungsproblem" introduced the Turing machine, a formal model of computation that demonstrated how mechanical processes could simulate any algorithmic behavior, influencing later ideas of agents as rule-following automata capable of decision-making in dynamic environments. During the 1950s and 1960s, these ideas manifested in pioneering AI projects that prototyped agent-like behaviors. John McCarthy's 1959 paper "Programs with Common Sense" proposed the "advice-taker," a hypothetical program designed to solve problems by manipulating sentences in predicate calculus, incorporating external advice as declarative rules to deduce actions and improve performance without requiring knowledge of its internal structure—representing an early blueprint for reasoning agents that learn from instructions.^[9] This vision advanced through practical implementations, such as the Shakey robot project at Stanford Research Institute (1966–1972), where software integrated computer vision, natural language processing, and planning algorithms to enable the robot to navigate, map environments, and execute goal-directed tasks autonomously, marking the first demonstration of a mobile system combining perception, reasoning, and action.^[10] The 1970s and 1980s brought further refinement, particularly through distributed AI and knowledge-based systems, where the term "agent" gained traction for describing independent computational units. Carl Hewitt's 1977 paper "Viewing Control Structures as Patterns of Passing Messages" formalized the actor model, portraying computation as a network of autonomous actors that communicate via asynchronous messages to solve problems collaboratively, providing a theoretical foundation for distributed agents in concurrent and decentralized settings.^[11] Concurrently, initial applications in simulation and control systems emerged, exemplified by precursors to expert systems like DENDRAL (developed from 1965 onward), which used heuristic rules and domain knowledge to generate and evaluate hypotheses for molecular structures in mass spectrometry, illustrating early agentic capabilities in scientific problem-solving and automated decision-making.^[12] By the late 1980s, these elements coalesced in distributed AI research, where agents were conceptualized as proactive entities in multi-agent simulations for tasks like resource allocation and coordination.^[13]

Evolution in the Digital Age

The 1990s marked the rise of software agents alongside the expansion of the internet, transitioning theoretical concepts into practical digital implementations. Mobile agents, capable of migrating across networks to perform tasks autonomously, emerged as a key innovation, exemplified by Telescript, a programming language developed by General Magic and introduced in 1994 to enable secure, mobile code execution on remote devices.^[14] This period also saw the formalization of multi-agent systems (MAS) through standards like those from the Foundation for Intelligent Physical Agents (FIPA), established in 1996 to promote interoperability among heterogeneous agents in distributed environments. Seminal works, such as Michael Wooldridge and Nicholas R. Jennings' 1995 paper "Intelligent Agents: Theory and Practice," provided foundational frameworks for understanding agent autonomy, perception, and interaction, influencing subsequent research in distributed AI.^[5] In the 2000s, software agents integrated deeply with web services and the semantic web, enhancing discoverability and automation. Agents leveraged ontologies for knowledge representation, particularly with the Web Ontology Language (OWL), standardized by the W3C in 2004, which enabled semantic reasoning in web-based agent interactions. This era also featured OWL-S, an extension for describing web services semantically, allowing agents to compose and invoke services dynamically. Embedded agents appeared in consumer devices, such as the iRobot Roomba vacuum cleaner launched in 2002, which used reactive software agents for navigation and obstacle avoidance in physical environments. The 2010s witnessed an AI boom that propelled software agents through advances in machine learning, particularly reinforcement learning (RL). DeepMind's AlphaGo, released in 2016, exemplified RL agents by mastering the game of Go through self-play and neural network policies, achieving superhuman performance and demonstrating scalable decision-making in complex domains. Concurrently, conversational agents gained prominence with Apple's Siri in 2011, an early voice-activated assistant integrating natural language processing for task delegation and user interaction. These developments shifted agents from isolated systems to interactive, learning entities embedded in everyday applications. By the 2020s, large language models (LLMs) transformed software agents into versatile, goal-oriented systems, with frameworks like LangChain (introduced in 2022) enabling modular agent construction using LLM chains for planning and execution.^[15] Auto-GPT, launched in 2023, represented an early LLM-powered autonomous agent capable of iterative task decomposition without constant human input, sparking widespread experimentation in agentic AI. Agents proliferated in cloud computing environments, leveraging scalable infrastructure for distributed operations, as outlined in AWS guidance on agent evolution.^[16] Emerging 2025 trends include decentralized agents integrated with blockchain for secure, peer-to-peer coordination, enhancing autonomy in Web3 ecosystems.^[17] This period also emphasized shifts toward hybrid human-AI agents, where agents augment human decision-making through collaborative interfaces, as explored in recent multi-agent system reviews.^[18]

Conceptual Distinctions

Agents vs. Traditional Programs and Objects

Software agents differ fundamentally from traditional programs in their level of autonomy and interaction with the environment. Traditional programs, often implemented as deterministic scripts, execute a fixed sequence of instructions in response to direct user input or predefined triggers, lacking the ability to perceive or adapt to changes independently.^[19] In contrast, software agents operate autonomously toward user-defined goals, continuously monitoring their environment, making decisions, and adjusting behaviors without constant human intervention.^[20] This goal-oriented nature enables agents to handle dynamic scenarios, such as proactively detecting and responding to unexpected events, whereas traditional programs remain reactive and rigid.^[2] For instance, a traditional program for tax calculation processes input data according to hardcoded rules to produce a one-time output, requiring manual invocation each time changes occur.^[21] A software agent, however, might continuously monitor financial accounts, analyze spending patterns in real time, and adapt tax strategies—such as suggesting deductions or alerting to compliance risks—based on evolving data and goals.^[22] Compared to objects in object-oriented programming (OOP), software agents extend beyond mere encapsulation and inheritance by exhibiting proactive, dynamic behaviors in situated environments. OOP objects are passive entities that respond predictably to method calls through predefined interfaces, maintaining state within their boundaries but without inherent initiative or environmental awareness.^[23] Agents, by contrast, possess their own thread of control, enabling them to initiate actions, engage in ongoing interactions via rich communication protocols, and even migrate across systems to pursue objectives.^[23] Furthermore, agents can learn and modify their behavior at the instance level in response to environmental feedback, a capability not native to static OOP objects.^[23] This distinction is rooted in the theoretical foundation of agents as "situated" entities, which emphasize real-time, improvisational activity within a dynamic context rather than abstract, plan-based computation characteristic of traditional programs or objects.^[24] In the seminal Pengi system, Agre and Chapman (1987) demonstrated how situated agents achieve complex, adaptive performance through simple, environment-coupled mechanisms, bypassing the need for explicit internal models or rigid scripts.^[24]

Agents vs. Expert Systems and Broader AI Agents

Software agents differ from expert systems primarily in their degree of autonomy, adaptability, and interaction with dynamic environments. Expert systems, such as MYCIN developed in 1976, are rule-based programs designed for narrow, domain-specific problem-solving, relying on static knowledge bases and human-provided inputs without proactive sensing or adaptation to changing conditions.^[25] In contrast, software agents operate autonomously in open environments, sensing their surroundings, pursuing goals over time, and adjusting behaviors based on feedback, which enables proactivity and flexibility beyond the rigid, synchronous reasoning of expert systems.^[26]^[27] While broader AI agents encompass a wide range of entities—including biological systems, robotic embodiments like Rodney Brooks' Herbert, and artificial life forms—software agents represent a specific subset confined to computational implementations in code, emphasizing digital execution without physical interaction.^[28] This distinction highlights the digital-only focus of software agents, which are situated within virtual environments and act through software mechanisms, whereas AI agents may integrate hardware for real-world sensing and actuation.^[28] Franklin and Graesser (1996) define autonomous agents broadly as "a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda," positioning software agents as a key implementation category within this taxonomy.^[28] Along the intelligence spectrum, software agents range from simple, rule-based variants—such as reactive thermostats that respond to immediate stimuli without planning—to more advanced, learning-enabled ones that incorporate adaptation and goal-oriented decision-making, though they do not presuppose full AI rationality or human-like cognition.^[28] This gradation allows software agents to bridge basic automation with intelligent behavior, distinguishing them from the purely declarative, non-learning nature of expert systems while remaining a focused subset of the diverse AI agent landscape.^[28]^[27]

Architectures and Design Principles

Agent Architectures

Software agent architectures provide the structural frameworks for designing systems that perceive their environment, make decisions, and act autonomously to achieve goals. These paradigms range from simple reactive models to complex deliberative and hybrid designs, each balancing reactivity, planning, and adaptability based on computational constraints and application needs.^[6] Reactive architectures emphasize direct mapping from environmental stimuli to responses, avoiding explicit internal representations or planning to enable fast, real-time operation in dynamic settings. A seminal example is the subsumption architecture, introduced by Rodney Brooks in 1986, which organizes behaviors into layered finite-state machines where higher layers can suppress (subsumption) lower ones to prioritize urgent actions, such as obstacle avoidance in mobile robots. This approach suits embedded systems requiring robustness without deliberation, as demonstrated in early robotic implementations.^[29] Deliberative architectures, in contrast, incorporate symbolic reasoning and planning to model the agent's internal state and goals explicitly. The Belief-Desire-Intention (BDI) model, developed by Rao and Georgeff in the early 1990s, structures agents around three core components: beliefs representing the agent's knowledge of the world, desires as potential goals or options, and intentions as committed plans derived from desires via reasoning. This rational framework enables agents to deliberate over actions by filtering desires into feasible intentions, supporting applications in complex, goal-oriented environments like automated planning systems.^[30] Hybrid architectures integrate reactive and deliberative elements to leverage the strengths of both, often through layered designs that allow low-level reactivity for immediate responses alongside higher-level planning. The InteRRaP architecture, proposed by Müller in the mid-1990s, exemplifies this by dividing agent functionality into three layers—social (interaction with others), reactive (behavior patterns), and planning (deliberative goal achievement)—enabling flexible control in multi-agent scenarios such as cooperative problem-solving.^[31] In BDI architectures, decision-making follows rational agent theory, where the agent selects the action a that maximizes expected utility:

U(a) = \arg\max_a \sum_{s} P(s \mid a) \cdot U(s),

with U(s) denoting the utility of state s and P(s \mid a) the probability of reaching s given action a. This equation derives from the principle that a rational agent, facing uncertainty, chooses actions to optimize performance measures over possible outcomes, as formalized in foundational AI texts; the summation computes the expected value, ensuring intentions commit to high-reward plans.^[6] Modern extensions to these architectures in the 2020s increasingly incorporate neural networks for enhanced perception and learning, particularly in handling unstructured data like natural language or images, while retaining core deliberative or hybrid structures for decision-making. For instance, neural components augment belief updates in BDI agents by processing sensory inputs via deep learning models, improving adaptability in real-world software agents integrated with large language models.

Frameworks and Implementation Models

Software agents are often developed using agent-oriented programming (AOP) languages that emphasize concepts like autonomy, reactivity, and social ability. One seminal AOP language is AgentSpeak, introduced in the mid-1990s specifically for implementing belief-desire-intention (BDI) agents. AgentSpeak(L), formalized in 1995, provides a logical, executable framework where agents are programmed through beliefs, plans, and events, enabling declarative specifications of agent behavior in multi-agent systems.^[32] This language has influenced subsequent extensions, such as AgentSpeak(ER) for encapsulation and AgentSpeak(PL) for probabilistic beliefs via Bayesian networks, facilitating robust decision-making under uncertainty.^[33] Practical implementation of software agents relies on dedicated platforms that provide middleware for communication, coordination, and deployment. The Java Agent DEvelopment Framework (JADE), released in 2000, is a widely adopted open-source platform for building FIPA-compliant multi-agent systems (MAS), offering tools for agent lifecycle management, message passing, and directory services to ensure interoperability.^[34] For Python-based development, SPADE (Smart Python Agent Development Environment), emerging in the early 2010s, leverages XMPP for instant messaging protocols, allowing agents to interact seamlessly with both other agents and human users in distributed environments.^[35] More recently, LangGraph, introduced in 2024 by LangChain and reaching its stable 1.0 version in October 2025, serves as a low-level orchestration framework tailored for large language model (LLM)-based agents, enabling the construction of stateful, graph-structured workflows with features like durable state persistence and human-in-the-loop support for complex, resilient agent applications.^[36]^[37] Key interaction models in software agent development include protocols for negotiation and task allocation in MAS. The Contract Net Protocol, originally proposed in 1980 by Reid G. Smith, formalizes a high-level communication mechanism where a manager agent announces tasks, and potential contractor agents bid through negotiation to secure contracts, promoting efficient distributed problem-solving.^[38] This protocol, evolved from earlier 1970s ideas in distributed AI, remains foundational for agent coordination, with implementations in various frameworks to handle dynamic resource allocation.^[39] Implementing software agents in distributed environments presents challenges, particularly scalability, where coordinating numerous agents across networks requires robust middleware to manage latency, fault tolerance, and synchronization. The Robot Operating System (ROS), initiated in 2007, addresses these by providing a flexible middleware suite for agent-like robotic components, facilitating message passing, service discovery, and modular coordination in real-time systems.^[40] Such middleware mitigates bottlenecks in large-scale deployments, ensuring agents can operate reliably in heterogeneous settings. As of 2025, advancements in agentic workflows emphasize enhanced memory mechanisms through integration with vector databases, enabling agents to store and retrieve semantic embeddings of past interactions for improved context awareness and long-term reasoning. Frameworks like LangGraph now commonly incorporate vector stores such as Pinecone or Weaviate to persist agent states, allowing scalable memory retrieval in LLM-driven applications without overwhelming computational resources.^[41] This integration has become standard for building adaptive, memory-augmented agents that maintain coherence over extended interactions.

Types and Examples

Autonomous and Personal Agents

Autonomous and personal agents represent a class of software agents designed to operate independently on behalf of individual users, handling routine tasks and providing personalized support without constant human oversight. These agents exhibit high degrees of autonomy by perceiving user needs, making decisions, and executing actions in dynamic environments, often integrating natural language processing and machine learning to adapt over time. Unlike reactive systems that respond only to explicit inputs, personal agents proactively anticipate requirements, such as by monitoring calendars or preferences to initiate actions unprompted.^[42] A prominent example of personal agents is virtual assistants like Google Assistant, which was launched in May 2016 as an evolution from earlier Google voice technologies, initially integrated into devices like the Google Pixel smartphone and later expanded to smart speakers and wearables. By the 2020s, these assistants evolved into multimodal systems capable of processing voice, text, visual, and even gesture inputs, enabling more intuitive interactions such as analyzing images for recommendations or combining audio queries with on-screen visuals for complex tasks. This progression has allowed agents to support diverse personal activities, from managing daily schedules to providing contextual advice, enhancing user productivity while maintaining a focus on individual utility.^[43]^[44] Buyer and shopping agents exemplify autonomous personal agents in e-commerce, automating price comparisons and negotiations to optimize purchases for users. One early instance is PriceGrabber, founded in 1999 as a price-comparison platform that aggregates offers from multiple retailers, allowing users to delegate search tasks for the best deals without manual effort. In modern contexts, AI-driven shopping agents leverage reinforcement learning to negotiate dynamically, as demonstrated in experimental models where buyer agents learn to propose counteroffers and adapt strategies to secure lower prices in simulated e-commerce scenarios, balancing user budgets with seller constraints. These agents autonomously evaluate market data, predict optimal timing, and execute transactions, reducing cognitive load for consumers.^[45]^[46] Key autonomy features in personal agents include task delegation capabilities, where users assign high-level goals, and the agent breaks them into subtasks, such as scheduling meetings by checking availability, sending invitations, and sending reminders without further input. Recommendation systems further illustrate this by proactively suggesting actions based on historical data, like curating shopping lists or travel itineraries aligned with user preferences, learned through ongoing interaction. These features enable seamless integration into daily life, with agents handling interruptions or changes autonomously to ensure reliability.^[42]^[47] IBM's Watson serves as an early precursor to personal agents, debuting in 2011 through its victory on the Jeopardy! quiz show, which showcased its natural language understanding and question-answering prowess, laying groundwork for assistant-like applications in healthcare and customer service. By 2025, Watson variants within the watsonx platform have incorporated advanced privacy-focused features, such as encrypted prompt processing and data isolation in IBM Cloud environments, ensuring user interactions remain secure without third-party exposure, thus addressing ethical concerns in personal agent deployment. This evolution highlights how foundational AI systems have matured into privacy-centric tools for individual empowerment.^[48]^[49]

Collaborative and Communication Agents

Collaborative and communication agents are software entities engineered to interact and coordinate with other agents or systems in networked or distributed settings, facilitating collective intelligence for complex tasks that exceed the scope of solitary agents. These agents emphasize interoperability through standardized messaging, enabling negotiation, information exchange, and joint decision-making in dynamic environments. A foundational protocol for such communication is the Agent Communication Language (ACL), specified by the Foundation for Intelligent Physical Agents (FIPA) in 1997, which supports message passing via performative acts like inform, request, and propose to ensure semantic interoperability among heterogeneous agents.^[50] This language underpins agent dialogues by defining message structure, including sender, receiver, content, and ontology, allowing agents to perform speech acts that align with their social abilities in multi-agent interactions. In multi-agent systems (MAS), coordination protocols enable agents to synchronize actions for applications such as resource auctions and traffic management. For instance, auction-based MAS allow agents to bid competitively for shared resources, as demonstrated in models where vehicles or traffic signals participate in Vickrey-style auctions to resolve intersection conflicts and minimize delays.^[51] Similarly, in traffic management, multi-agent reinforcement learning frameworks coordinate distributed agents controlling traffic lights, achieving improvements in throughput by learning cooperative policies without central oversight.^[52] Hierarchical coordination emerged in the 1990s through holonic MAS, inspired by the Holonic Manufacturing Systems (HMS) project initiated in 1994, where holons—autonomous, cooperative subunits—form recursive hierarchies to manage manufacturing workflows, such as order holons delegating tasks to resource holons for flexible production scheduling.^[53] Practical examples illustrate these principles in action. Email filtering agents, exemplified by SpamAssassin released in 2001, operate in distributed setups where multiple agents exchange metadata on email patterns to collaboratively detect and quarantine spam, integrating rule-based heuristics with network-shared blacklists.^[54]^[55] In swarm robotics simulations, agents communicate locally via infrared or radio signals to achieve emergent behaviors; the Kilobot platform, developed for scalable collectives, enables hundreds of simple agents to self-organize for tasks like pattern formation through neighbor-to-neighbor messaging.^[56] By 2025, advancements in federated learning have introduced privacy-preserving collaborative agents within MAS, where distributed agents train shared models without exchanging raw data, as in the Helmsman framework that uses multi-agent synthesis to autonomously configure federated systems for edge computing applications.^[57] These developments enhance scalability and security in decentralized environments, building on ACL-like protocols for secure aggregation of model updates.

Specialized Agents (Monitoring, Security, and Development)

Specialized agents in monitoring, security, and development domains leverage autonomy to perform targeted tasks, such as real-time surveillance, threat mitigation, and code assistance, enhancing system reliability without constant human oversight. These agents often integrate machine learning for adaptive decision-making, distinguishing them from general-purpose tools by their focus on niche operational efficiency. Monitoring and predictive agents detect anomalies and forecast issues in dynamic environments, particularly in network traffic and Internet of Things (IoT) ecosystems. For instance, Snort, an open-source network intrusion detection system (IDS) developed in 1998, uses rule-based signatures to monitor packet streams for malicious patterns, alerting administrators to potential breaches in real time.^[58] In the 2020s, predictive maintenance agents in IoT settings analyze sensor data to anticipate equipment failures, employing AI models to predict remaining useful life and schedule interventions proactively.^[59] These agents, structured around core modules for data collection, analysis, and notification, reduce downtime in industrial applications through early anomaly detection.^[60] Security agents extend monitoring capabilities with proactive defenses, including honeypots and adaptive firewalls that evolve against emerging threats. Honeypots serve as decoy systems to attract and study attackers, logging interactions to refine intrusion detection without risking production assets, as seen in deployments that minimize false positives compared to traditional IDS.^[61] Adaptive firewalls incorporate machine learning to dynamically adjust rules based on traffic patterns, classifying anomalies with accuracies exceeding 95% using models like Random Forest for real-time threat blocking.^[62] Data-mining agents, integrated with toolkits like WEKA since the 1990s, enhance security by discovering hidden patterns in logs, such as unusual access behaviors, through clustering and classification algorithms that support forensic analysis.^[63]^[64] The rise of self-healing security agents post-2020 has addressed escalating cyber threats, with AI-driven systems autonomously detecting, isolating, and repairing vulnerabilities, such as patching exploited endpoints without manual intervention.^[65] These agents use reinforcement learning to adapt responses, improving resilience in enterprise networks amid incidents like ransomware surges.^[66] Development agents automate software creation and validation, streamlining workflows in continuous integration/continuous deployment (CI/CD) pipelines. GitHub Copilot, released in 2021 by GitHub and OpenAI, acts as an AI pair programmer, suggesting code completions and functions based on context, boosting developer productivity in supported editors.^[67] Auto-testing agents in CI/CD, such as those leveraging AI for test case generation and execution, integrate with pipelines to run parallel checks and self-correct failures, reducing manual debugging in agile environments.^[68] By analyzing code changes and historical data, these agents ensure higher test coverage and faster release cycles, with frameworks enabling autonomous iteration on defects.^[69]

Applications and Societal Impacts

Organizational and Economic Effects

Software agents have profoundly influenced organizational structures by automating repetitive workflows, enabling businesses to reallocate human resources toward higher-value activities. Robotic Process Automation (RPA) agents, a prominent category of software agents, mimic human actions to handle rule-based tasks such as data entry and invoice processing, thereby reducing manual labor intensity. According to a 2024 Gartner report, 30% of enterprises are projected to automate more than half of their network activities by 2026, up from under 10% in mid-2023, demonstrating the accelerating adoption of such agents for operational efficiency.^[70] This shift allows organizations to streamline processes across departments, fostering agile decision-making and reducing errors in complex environments like finance and HR. Economically, software agents drive substantial cost savings while posing challenges related to workforce displacement. In e-commerce, AI agents manage personalized recommendations and customer interactions, optimizing inventory and marketing efforts to lower operational expenses. McKinsey analysis indicates that agentic AI in retail could enable autonomous transactions and hyperpersonalization, yielding significant productivity gains in sales operations.^[71] For instance, these agents can reduce logistics costs by enhancing demand forecasting and route optimization, with broader AI implementations projected to save the global logistics sector up to $1.5 trillion annually by 2030.^[72] However, in routine sectors such as administrative support and manufacturing, the deployment of agents contributes to job displacement, as automation substitutes for labor in predictable tasks; Goldman Sachs estimates that AI could affect nearly 300 million full-time jobs globally, necessitating reskilling initiatives to mitigate economic inequality.^[73] Adoption trends highlight the integration of multi-agent systems in supply chains, where coordinated agents enhance resilience and efficiency. These systems involve multiple specialized agents collaborating to manage inventory, predict disruptions, and optimize logistics in real time. A notable example is Amazon's use of multi-agent reinforcement learning in warehouse sortation centers, where agents allocate resources across hundreds of robotic units to minimize delays and reduce unsorted packages, improving throughput in high-volume fulfillment operations.^[74] Such implementations have become standard in retail and manufacturing. Looking ahead, the economic impact of agentic software agents is poised for exponential growth. PwC projects that AI, including advanced agent systems, could add up to $15.7 trillion to global GDP by 2030 through productivity enhancements and new consumption patterns, equivalent to a 14% increase over baseline forecasts.^[75] This contribution underscores the transformative potential of agents in reshaping economic models, from cost-efficient automation to innovative business ecosystems. As of 2025, reports indicate growing enterprise adoption of agentic AI, with investments accelerating in multi-agent frameworks for supply chain optimization. Software agents have significantly influenced work contentment by automating repetitive tasks, thereby allowing employees to focus on more creative and fulfilling aspects of their roles. A study involving administrative and HR professionals found that generative AI tools, which function as software agents, reduced time spent on routine activities, leading to increased reported enjoyment of these tasks.^[76] However, this automation has also raised concerns about deskilling, where over-reliance on agents diminishes workers' manual skills and problem-solving abilities in foundational tasks, potentially leading to reduced job autonomy and long-term dissatisfaction.^[77] Culturally, software agents, particularly AI companions, are reshaping social norms by integrating into everyday interactions and media consumption. For instance, AI companions used in social media and entertainment platforms can normalize reliance on non-human entities for emotional support, altering expectations around human relationships and potentially fostering isolation if they substitute genuine social bonds. This dependency culture extends to daily life, where autonomous personal agents handle routine decisions like scheduling or recommendations, creating a broader societal shift toward convenience-driven behaviors that may erode independent decision-making skills over time.^[78]^[79] Ethical challenges posed by software agents include bias in decision-making processes, which can perpetuate discrimination. In e-commerce, recommendation agents have been observed to favor products based on skewed training data, such as suggesting luxury items predominantly to higher-income demographics while limiting options for underrepresented groups, thereby reinforcing socioeconomic divides. Privacy erosion arises from agents' continuous monitoring of user behaviors to personalize services, often without adequate consent, leading to unauthorized data aggregation and heightened surveillance risks. Additionally, accountability remains elusive in cases of autonomous failures, such as an agent making erroneous financial decisions, where liability is unclear between developers, deployers, and users, complicating redress for affected parties.^[80]^[81]^[82]^[83] As of 2025, regulatory efforts address these implications through frameworks like the EU AI Act, enacted in 2024, which classifies certain software agents as high-risk systems and mandates transparency measures, including disclosure of operational limitations and risk assessments, to ensure ethical deployment and user trust.^[84]

References

[1]
[PDF] Software Agents
for a more in depth analysis ...
[2]
Software Agents: An Overview - Systems and Computer Engineering
This paper largely reviews software agents, and it also contains some strong opinions that are not necessarily widely accepted by the agent community.Missing: authoritative | Show results with:authoritative<|control11|><|separator|>
[3]
[PDF] Software Agents
All of these three classes of agents have “Agenthood”: Page 3. The Key Characteristics of “Agenthood”: (Wooldridge and Jennings, 1995):. • Autonomy: agents ...
[4]
Artificial Intelligence: A Modern Approach, 4th US ed.
Artificial Intelligence: A Modern Approach, 4th US ed. by Stuart Russell and Peter Norvig. The authoritative, most-used AI textbook, adopted by over 1500 ...Missing: definition | Show results with:definition
[5]
[PDF] Intelligent agents: theory and practice
Finally, agent languages are software systems for programming and experimenting with agents; these languages ... (1990), Jennings (1993a), Wooldridge (1994) and.
[6]
[PDF] 2 INTELLIGENT AGENTS - People @EECS
They will all have the same skeleton, namely, accepting percepts from an environment and generating actions. The early versions of agent programs will have a ...
[7]
Rationality and intelligence - ScienceDirect.com
This paper outlines a gradual evolution in the formal conception of rationality that brings it closer to our informal conception of intelligence.
[8]
Cybernetics or Control and Communication in the Animal and the ...
Norbert Wiener, Cybernetics, or Control and Communication in the. Animal and the Machine (Cambridge, MA: MIT Press, 1948), 42. 5. Wiener, Cybernetics, 162.
[9]
[PDF] PROGRAMS WITH COMMON SENSE - Formal Reasoning Group
It was my impression that Dr. McCarthy's advice taker was meant to be able, among other things, to arrive at a certain conclusion from appropriate premises by ...Missing: agent | Show results with:agent
[10]
Milestones:SHAKEY: The World's First Mobile Intelligent Robot, 1972
Feb 12, 2024 · Shakey's control software was structured as a multi-level hierarchy with physical actions at the lowest levels, autonomous planning in a middle ...
[11]
https://doi.org/10.1016/0004-3702(77)90033-9
[12]
[PDF] DENDRAL: a case study of the first expert system for scientific ... - MIT
The DENDRAL. Project was one of the first large-scale programs to embody the strategy of using detailed, task-specific knowledge about a problem domain as a ...
[13]
Agentic Discovery: Closing the Loop With Cooperative Agents
However, popular use of the term “agent” emerged throughout the 1980s within distributed AI. Then, and continuing into the 1990s, AI was often regarded as ...
[14]
Software Agents for Future Communication Systems
emerging standards defined by the Foundation for Intelligent Physical Agents (FIPA) ... (1996) 'Telescript Technology: Mobile Agents', General magic White Paper.<|separator|>
[15]
LangChain
LangChain provides the engineering platform and open source frameworks developers use to build, test, and deploy reliable AI agents.LangChain overview · Agents · LangChain Academy · Build agents faster, your wayMissing: 2023 2022 decentralized blockchain 2025 hybrid
[16]
The evolution of software agents - AWS Prescriptive Guidance
How traditional software agents evolved from the 1950s concept of autonomy through acting agents to LLMs and agentic AI.
[17]
LLM-Based Multi-Agent Systems for Software Engineering
This article explores the transformative potential of integrating Large Language Models into Multi-Agent (LMA) systems for addressing complex challenges in ...
[18]
IEEE Power & Energy Society Multi-Agent Systems Working Group
Wooldridge describes this pro-activeness as an agent's ability to “take the initiative.” Social ability: intelligent agents are able to interact with other ...
[19]
Agents that Buy and Sell - Communications of the ACM
Mar 1, 1999 · Unlike so-called traditional software, software agents are personalized, continuously running, and semiautonomous [1].
[20]
Can AI Do Business Taxes? Benefits, Limitations, and Current Uses
Examples. Intuit Assist, an AI assistant in certain TurboTax and Quickbooks software packages, uses traditional and generative AI that answers tax questions, ...
[21]
The Rise of AI Agents in Finance: Preparing Controllers for ...
Sep 16, 2025 · Monitor Finance Activity: AI agents can watch data in real time, flagging anomalies or triggering workflows without waiting for a human prompt ...
[22]
Objects and Agents Compared - JOT: Journal of Object Technology
This paper discusses some of the differences and similarities between agents and objects and lets you decide which viewpoint you want to choose.
[23]
Pengi: An Implementation of a Theory of Activity - AAAI
Pengi: An Implementation of a Theory of Activity. February 1, 2023 ... Download PDF. Abstract: AI has generally interpreted the organized ...
[24]
Mycin: A Knowledge-Based Computer Program Applied to Infectious ...
Mycin: A Knowledge-Based Computer Program Applied to Infectious Diseases. Edward H Shortliffe. Edward H Shortliffe. Find articles by Edward H Shortliffe.
[25]
[PDF] a comparison between expert systems and - SciELO Colombia
Expert systems belong to a branch of artificial intelligence named knowledge engineering. An expert system can be understood as a software mechanism used to ...
[26]
[PDF] 1 Applications of Intelligent Agents
Many agent architectures have been developed by the intelligent agents community, with many different properties (Wooldridge and Jennings, 1995). At one ...
[27]
[PDF] Is it an Agent, or just a Program?: A Taxonomy for Autonomous Agents
An autonomous agent is a software agent that operates without direct human intervention, has control over its actions, and is distinct from just any program.
[28]
[PDF] A Robust Layered Control System for a Mobile Robot
We describe a new architecture for controlling mobile robots. Layers of control system are built to let the robot operate at increasing levels of competence.
[29]
[PDF] BDI Agents: From Theory to Practice
The BDI architecture draws its inspiration from the philosophical theories of Bratman (Bratman 1987), who argues that intentions play a significant and dis-.
[30]
[PDF] The Agent Architecture InteRRaP: Concept and Application
In the rst part of this report, InteRRaP, an agent architecture for multi-agent systems is presented. The basic idea is to combine the use of patterns of ...
[31]
BDI agents speak out in a logical computable language - SpringerLink
Jun 25, 2005 · In this paper, we provide an alternative formalization of BDI agents by providing an operational and proof-theoretic semantics of a language AgentSpeak(L).
[32]
AgentSpeak(PL): A New Programming Language for BDI Agents ...
AgentSpeak(PL) integrates the concept of probabilistic beliefs through the use of Bayesian Networks, to core BDI programming concepts. The language is ...
[33]
JADE: a FIPA2000 compliant agent development environment
This paper describes the main features of the JADE system and introduces some of the most important projects based on JADE software. Formats available. You can ...
[34]
javipalanca/spade: Smart Python Agent Development Environment
A multi-agent systems platform written in Python and based on instant messaging (XMPP). Develop agents that can chat both with other agents and humans.
[35]
LangGraph - LangChain
from conversational agents, complex task automation, to custom LLM-backed ...
[36]
[PDF] The Contract Net Protocol: High-Level Communication and Control ...
The contract net protocol specifies communication and control for distributed problem solvers, using negotiation between nodes to distribute tasks and control.
[37]
The Evolution of the Contract Net Protocol - ResearchGate
Aug 7, 2025 · PDF | Contracts are a powerful co-ordination mechanism in distributed systems. The contract net protocol has been applied since about 1980.Missing: 1970s | Show results with:1970s
[38]
[PDF] ROS: an open-source Robot Operating System - Stanford AI Lab
In this paper, we discuss how ROS relates to existing robot software frameworks, and briefly overview some of the available application software which uses ROS.Missing: agent coordination
[39]
What Is AI Agent Memory? | IBM
By integrating vector databases, agentic systems can efficiently store embeddings of previous interactions, enabling contextual recall. This is useful for AI- ...
[40]
What are Autonomous Agents? A Complete Guide - Salesforce
Autonomous agents are an advanced form of AI that can independently execute a series of tasks, and learn as they go. Here's what that means for your business.
[41]
What is Google Assistant? Here's the guide you need to get started
Mar 19, 2023 · Google Now persisted for years, before eventually being absorbed into Google Assistant when it launched on the first Google Pixel. Google ...
[42]
Virtual Assistants: A Review of the Next Frontier in AI Interaction
May 16, 2025 · This review seeks to consolidate disparate research by offering a novel comprehensive taxonomy of virtual assistants, as well as smart speakers.
[43]
PriceGrabber.com 2025 Company Profile - PitchBook
PriceGrabber.com was founded in 1999. Where is PriceGrabber.com headquartered? PriceGrabber.com is headquartered in Los Angeles, CA. What is the size ...
[44]
The Automated but Risky Game: Modeling and Benchmarking Agent ...
Sep 20, 2025 · Inspired by real-world shopping and sales workflows, we design an experimental setting where a buyer agent attempts to negotiate lower prices ...
[45]
5 AI agents that automate mundane tasks effortlessly - Glean
Aug 6, 2025 · The Glean Team | 5 AI agents that automate mundane tasks effortlessly save you hours by handling email triage, data entry, scheduling and ...
[46]
IBM Watson
IBM Watson technology became available as a development platform in the cloud. The move spurred innovation and fueled a new ecosystem of entrepreneurial ...IBM watsonx · IBM Garage · Explore watsonx.ai · Explore watsonx.dataMissing: launch | Show results with:launch
[47]
Security and privacy for foundation models — Docs | IBM watsonx
Oct 23, 2025 · The following table summarizes the privacy and security of your foundation model work in watsonx. Table 1. Privacy and security in watsonx.ai ...Missing: variants | Show results with:variants
[48]
[PDF] FIPA 1997 part 2: Agent Communication Language
Oct 10, 1997 · The FIPA 97 specification is the first output of the Foundation for Intelligent Physical Agents. It provides specification of. 160 basic agent ...Missing: paper | Show results with:paper
[49]
[PDF] A Multi-agent Auction-based Approach for Modeling of Signalized ...
Abstract.This paper shows how the traffic interactions of an intersection can be regulated by means of auction theory, multi-agent systems and machine.
[50]
[PDF] Multiagent Reinforcement Learning for Urban Traffic Control using ...
This paper extends the reinforcement learning approach to traffic control by using cooperative learning and explicit coordination among agents. We make the ...
[51]
[PDF] HOLONIC MANUFACTURING SYSTEMS: - Holobloc
Dec 1, 1994 · Holonic Manufacturing Systems (HMS) was one of the six test cases. The HMS. Consortium consisted of 31 Partners from all regions in the IMS ...
[52]
Apache SpamAssassin: Welcome
Apache SpamAssassin is the #1 Open Source anti-spam platform giving system administrators a filter to classify email and block spam (unsolicited bulk email).Downloads · Documentation · Mail::SpamAssassin · News and Announcements<|separator|>
[53]
https://holobloc.com/papers/hannover.pdf
[54]
Kilobot: A low cost robot with scalable operations designed for ...
This paper presents Kilobot, an open-source, low cost robot designed to make testing collective algorithms on hundreds or thousands of robots accessible to ...Missing: original | Show results with:original
[55]
Autonomous Synthesis of Federated Learning Systems via Multi ...
Oct 16, 2025 · [Submitted on 16 Oct 2025]. Title:Helmsman: Autonomous Synthesis of Federated Learning Systems via Multi-Agent Collaboration. Authors:Haoyuan ...
[56]
[PDF] Snort – Lightweight Intrusion Detection for Networks - USENIX
Martin Roesch is a Network Security Engineer with Stanford Telecommunications Inc. He holds a. B.S. in Computer Engineering from Clarkson Univer- sity. He has ...Missing: history | Show results with:history
[57]
Artificial Intelligence Agent-Enabled Predictive Maintenance - MDPI
In the proposed AI agent-enabled predictive maintenance framework, the system is structured around four core agents with clearly defined responsibilities: the ...Missing: 2020s | Show results with:2020s
[58]
(PDF) Intelligent Agents for Enhanced Predictive Maintenance and ...
Sep 21, 2025 · The agents are developed to select from historical failure data and output operating parameters to predict remaining useful life (RUL) and ...
[59]
What is a honeypot? How honeypots help security - Kaspersky
Honeypots have a low false positive rate. That's in stark contrast to traditional intrusion-detection systems (IDS) which can produce a high level of false ...
[60]
(PDF) ML-Powered Firewall for Adaptive Threat Detection and Real ...
Aug 6, 2025 · Among the several machine learning models that were employed and evaluated Random Forest obtained an accuracy rate of 95%, Extra Trees 94.85%, ...
[61]
[PDF] The WEKA Data Mining Software: An Update - SIGKDD
The WEKA project aims to provide a comprehensive collec- tion of machine learning algorithms and data preprocessing tools to researchers and practitioners alike ...Missing: agents 1990s
[62]
Data Mining and Agent Technology: a fruitful symbiosis
The following chapter aims to present such a unified and integrated methodology for a specific category of MAS. It takes all constraints and issues into account ...
[63]
(PDF) AI Driven Self-Healing Cybersecurity Systems with Agentic AI ...
Jun 13, 2025 · This study looks into the development of a self-healing security system which is integrated with Agentic AI, enabling it to have an ...Missing: rise | Show results with:rise
[64]
AI in Cyber Defense: The Rise of Self-Healing Systems
Mar 18, 2025 · AI-powered self-healing cybersecurity is transforming the industry by detecting, defending against, and repairing cyber threats without human intervention.Missing: 2020 | Show results with:2020
[65]
Introducing GitHub Copilot: your AI pair programmer
Jun 29, 2021 · Today, we are launching a technical preview of GitHub Copilot, a new AI pair programmer that helps you write better code.
[66]
CI/CD pipelines with agentic AI: How to create self ... - Elastic
Sep 30, 2025 · How our team introduced GenAI into CI pipelines to create self-correcting pull requests, automizing the update of hundreds of dependencies ...
[67]
[PDF] Autonomous AI Agents for End-to-End Data Engineering Pipelines ...
May 7, 2025 · The combination of autonomous AI agents with CI/CD pipelines introduces new paradigms in data engi- neering. By having an agent own certain ...
[68]
Gartner Says 30% of Enterprises Will Automate More Than Half of ...
Sep 18, 2024 · By 2026, 30% of enterprises will automate more than half of their network activities, an increase from under 10% in mid-2023, according to Gartner, Inc.
[69]
Agentic commerce: How agents are ushering in a new era - McKinsey
Oct 17, 2025 · Discover how agentic commerce uses AI shopping agents to transform retail with hyperpersonalized experiences, autonomous transactions, ...
[70]
AI Revolutionizing eCommerce Logistics: Efficiency, Cost Savings ...
Jul 27, 2023 · According to a report by McKinsey & Company, AI has the potential to save the global logistics industry up to $1.5 trillion by 2030. Companies ...
[71]
How Will AI Affect the Global Workforce? - Goldman Sachs
Aug 13, 2025 · AI-related innovation may cause near-term job displacement while also ultimately creating new opportunities elsewhere.Missing: cost | Show results with:cost
[72]
Multi-agent reinforcement learning for resource allocation in large ...
In this work, we propose a multi-agent reinforcement learning method to solve large-scale chute mapping problems with hundreds of agents (the destinations). To ...
[73]
Multi-Agent System: Top Industrial Applications in 2025 - [x]cube LABS
Aug 28, 2025 · According to Gartner, over 50% of enterprises are expected to adopt agent-based modeling by 2027 to enhance their decision-making capabilities.
[74]
[PDF] pwc-ai-analysis-sizing-the-prize-report.pdf
PwC research shows global GDP could be up to 14% higher in 2030 as a result of AI – the equivalent of an additional $15.7 trillion – making it the biggest ...
[75]
How Gen AI Can Make Work More Fulfilling
Jun 12, 2024 · Our earlier research has shown that employees who enjoy their work are about 50% less likely to look for a new job.
[76]
[PDF] Deskilling and upskilling with generative AI systems
Deskilling is a long-standing prediction of the use of information technology, raised anew by the increased capabilities of generative AI (GAI) systems.
[77]
Companionship in code: AI's role in the future of human connection
Jul 24, 2025 · This article builds on research into user experience and perceptions of AI companionship, informed by the author's everyday interactions with ...
[78]
Through the Chat Window and Into the Real World: Preparing for AI ...
If highly capable agents reach widespread use, users may become vulnerable to skill fade and dependency, agents may collude with one another in undesirable ways ...Missing: daily | Show results with:daily
[79]
Identifying AI Bias and AI Bias in the Retail Industry
For instance, an AI-powered recommendation engine might consistently suggest products that appeal to a narrow audience, neglecting the preferences or needs of ...Missing: agents | Show results with:agents
[80]
How AI Affects Product Recommendation Bias
Oct 30, 2024 · AI recommendations can be biased by profit margins, search costs, and the platform's fear of losing business, creating a gap between true ...
[81]
The Privacy and Security Risks of Autonomous AI Agents
Mar 14, 2025 · Organizations must prioritize robust AI governance, transparency, and security monitoring to balance automation benefits with data protection.
[82]
Who is responsible when AI acts autonomously & things go wrong?
May 15, 2025 · Even with autonomous AI, human or corporate actors remain accountable. Responsibility should be anchored to human decision-makers, not the AI ...
[83]
Article 13: Transparency and Provision of Information to Deployers
They must come with clear instructions, including information about the provider, the system's capabilities and limitations, and any potential risks.Missing: agents | Show results with:agents<|control11|><|separator|>