AI agent

An AI agent is an autonomous software system that perceives its environment and takes actions to achieve specified goals. Unlike passive AI systems that respond to single queries, AI agents operate continuously, make decisions, and execute multi-step plans with minimal human intervention.

The term encompasses systems from both the classical symbolic AI tradition and the newer neural/generative paradigm, though these operate on fundamentally different principles.

Definition

No universally accepted definition of "AI agent" exists, but several influential formulations capture key aspects:

Russell and Norvig (2020) define an agent broadly as "anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators." A rational agent is "one that acts so as to achieve the best outcome or, when there is uncertainty, the best expected outcome." ^[1]

Wooldridge and Jennings (1995) identify four key properties of intelligent agents: ^[2]

• Autonomy: Operating without direct human intervention and maintaining control over internal state
• Social ability: Interacting with other agents (human or artificial)
• Reactivity: Perceiving and responding to environmental changes
• Pro-activeness: Taking goal-directed initiative rather than merely reacting

Abou Ali et al. (2025) distinguish between an AI agent as "a self-contained autonomous system designed to accomplish a goal" that "operates primarily in isolation, though it may interact with tools and APIs," and agentic AI as "the broader field and architectural approach concerned with creating systems that exhibit agency." ^[3]

Terminology

Several related terms appear in the literature:

• Intelligent agent: A broader term from classical AI that includes non-AI systems (e.g., simple thermostats qualify under some definitions)
• Autonomous agent: Emphasizes independence from human control; often used interchangeably with AI agent
• Agentic AI: The field and paradigm concerned with creating systems that exhibit agency, often involving multi-agent coordination
• LLM agent: An AI agent built on a large language model as its core reasoning component

This article uses "AI agent" to refer to the autonomous systems themselves, regardless of underlying architecture.

Historical development

The development of AI agents can be understood through five overlapping eras: ^[3]

Symbolic AI era (1950s–1980s)

Early AI research focused on rule-based systems and explicit symbolic reasoning. Expert systems like MYCIN (medical diagnosis) and DENDRAL (chemical analysis) demonstrated that encoded human expertise could solve narrow problems. These systems were brittle—they failed on inputs outside their programmed rules—but provided explainable, verifiable behavior.

Machine learning era (1980s–2010s)

Statistical approaches replaced hand-crafted rules with learned patterns. Support vector machines, decision trees, and other algorithms enabled systems to generalize from training data. Reinforcement learning frameworks formalized how agents could learn optimal behavior through trial and error.

Deep learning era (2010s–present)

Neural networks with many layers achieved breakthrough performance on perception tasks. Convolutional neural networks transformed computer vision; recurrent networks enabled sequence processing. These advances gave agents new capabilities for processing unstructured sensory input.

Generative AI era (2014–present)

Generative adversarial networks (2014) and the transformer architecture (2017) enabled AI systems to produce novel content. Large language models like GPT and BERT demonstrated emergent capabilities in reasoning, planning, and natural language understanding that had not been explicitly programmed.

Agentic AI era (2022–present)

Beginning around 2022, researchers and developers began combining LLM capabilities with tool use, memory systems, and multi-step planning. Projects like AutoGPT demonstrated autonomous goal pursuit, while frameworks like AutoGen enabled multi-agent coordination. ^[4] This era is characterized by agents that can decompose complex tasks, use external tools, and operate over extended time horizons.

Architectural paradigms

Contemporary AI agents emerge from two distinct intellectual traditions that should not be conflated: ^[3]

Symbolic/classical paradigm

Classical agents use explicit representations of knowledge and algorithmic planning. Key frameworks include:

Markov Decision Processes (MDPs) and Partially Observable MDPs (POMDPs) provide mathematical frameworks for sequential decision-making under uncertainty. The agent maintains beliefs about world state and selects actions to maximize expected reward. ^[5]

BDI (Belief-Desire-Intention) architecture models agents in terms of mental states: beliefs about the world, desires (goals), and intentions (committed plans). This approach enables reasoning about agent behavior in human-understandable terms. ^[6]

Cognitive architectures like SOAR aim to model general intelligence through unified theories of cognition, incorporating learning, memory, and problem-solving in integrated systems. ^[7]

Symbolic agents offer verifiability, explainability, and predictable behavior, making them suitable for safety-critical applications. However, they struggle with unstructured data, require extensive hand-engineering, and scale poorly to complex environments.

Neural/generative paradigm

Neural agents use large language models as their reasoning core, with behavior emerging from training rather than explicit programming. Key characteristics include:

• Prompt-driven orchestration: Behavior is guided by natural language instructions rather than formal specifications
• Stochastic generation: Outputs are sampled from probability distributions, introducing inherent variability
• Tool use: LLMs invoke external APIs, databases, and code execution environments to extend their capabilities
• In-context learning: Agents adapt to new tasks through examples provided in their context window

Frameworks like LangChain, AutoGen, CrewAI, and Semantic Kernel provide scaffolding for building LLM-based agents. These systems excel at handling unstructured data, natural language interfaces, and novel situations, but suffer from opacity, hallucination, and difficulty in formal verification.

Hybrid approaches

Neuro-symbolic systems attempt to combine the strengths of both paradigms: neural components for perception and pattern recognition, symbolic components for reasoning and constraint satisfaction. This remains an active research direction with no dominant architecture.

Conceptual retrofitting—applying symbolic concepts like BDI or PPAR (Perceive-Plan-Act-Reflect) loops to describe neural systems—is common but potentially misleading, as LLM-based agents do not actually implement these architectures internally. ^[3]

Key capabilities

Modern AI agents typically exhibit some combination of:

Planning and reasoning

Agents decompose complex goals into subgoals and action sequences. Classical agents use search algorithms and formal planning; LLM agents often use chain-of-thought prompting or tree-of-thought exploration.

Memory

Beyond the immediate context window, agents may maintain:

• Short-term memory: Recent interactions and working state
• Long-term memory: Persistent knowledge, often stored in vector databases
• Episodic memory: Records of past experiences for learning and reference

Tool use

Agents extend their capabilities by invoking external systems: web search, code execution, database queries, API calls, and specialized software. Tool use transforms LLMs from text generators into systems that can affect the world. The [[Model Context Protocol]] (MCP) provides a standardized way for agents to discover and interact with external tools and data sources, enabling interoperability across different agent frameworks.

Multi-agent coordination

Complex tasks may involve multiple specialized agents collaborating: one agent for research, another for writing, another for code review. Frameworks like AutoGen provide protocols for agent communication and task delegation. ^[4]

Evaluation

Measuring agent performance remains challenging. Key benchmarks include:

AgentBench evaluates LLM agents across diverse environments including operating systems, databases, and web browsing, testing both individual capabilities and end-to-end task completion. ^[8]

GAIA focuses on real-world assistant tasks requiring multi-step reasoning, tool use, and web navigation. ^[9]

OSWorld tests multimodal agents on open-ended tasks in real computer environments across Ubuntu, Windows, and macOS. ^[10] When the benchmark launched in 2024, the best AI agents achieved only 12.24% success rates compared to human performance of 72.36%. Progress has been rapid: a preprint from Simular Research reported 69.9% success in October 2025, approaching human-level performance. ^[11][12]

Risks and governance

AI agents introduce distinct risks that differ by architectural paradigm: ^[13]

Symbolic system risks

Classical agents deployed in safety-critical domains (autonomous vehicles, medical devices, industrial control) can cause physical harm through specification errors, sensor failures, or edge cases not anticipated by designers. These systems are typically regulated through existing product safety frameworks, though governance has not kept pace with capability.

Neural system risks

LLM-based agents introduce novel challenges:

• Hallucination: Agents may act on false beliefs confidently stated
• Goal drift: Objectives may be misinterpreted or optimized in unintended ways
• Opacity: The reasoning behind actions is difficult to verify or predict
• Accountability gaps: When an LLM agent causes harm, attribution of responsibility is unclear

Governance frameworks

Practices for governing agentic AI systems remain underdeveloped. ^[14] Proposed approaches include:

• Constraining agent capabilities and access
• Requiring human approval for high-stakes actions
• Maintaining audit logs of agent decisions
• Implementing "meaningful human control" (though what this means differs substantially between paradigms)

Human agency implications

The deployment of AI agents raises fundamental questions about human agency:

Automation versus augmentation: Agents can either replace human decision-making or enhance human capabilities. The choice of which tasks to delegate affects skill development, employment, and the distribution of power.

Attention and autonomy: Agents that act on behalf of users may make choices those users would not have made themselves. Default settings and agent behaviors can shape human choices in ways that may not be transparent.

Accountability: When AI agents make consequential decisions, the question of who bears responsibility—developers, deployers, users, or the agents themselves—remains unresolved. This is particularly acute for neural systems where behavior cannot be fully predicted or explained.

Meaningful control: Maintaining human oversight becomes difficult as agents operate at speeds and scales that exceed human cognitive capacity, and as agent reasoning becomes less interpretable.