Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Autonomous Agent

nounAI Agents and Autonomous Systemsalso autonomous AI agent, autonomous software agent, self-directed agent

In one line

An autonomous agent is a system that senses its environment and acts on its own to achieve goals over time, making its own decisions about what to do next without continuous human control.

An autonomous agent is a system that senses its environment and acts on its own to achieve goals over time, making its own decisions about what to do next without continuous human control. The word autonomous stresses independence: the agent is not simply carrying out a script step by step under direct supervision but chooses actions based on what it perceives and what it is trying to do. Examples range from thermostats and warehouse robots to self-driving cars, space probes and software agents powered by large language models. The concept is central to the study of intelligent agents and to the current wave of AI agents.

Definition

In a widely cited 1996 taxonomy, Stan Franklin and Art Graesser described an autonomous agent as a system situated within an environment that senses it and acts on it over time, in pursuit of its own agenda and so as to affect what it senses in the future. Several features follow: the agent runs continuously or for long periods, it has goals, it perceives, it acts and its actions change the situation it will perceive next. Autonomy is a matter of degree, since agents depend on humans to set goals and limits.

Key properties

  • Situatedness: it exists in an environment, physical or digital, and responds to it.
  • Autonomy: it controls its own actions and internal state.
  • Reactivity: it responds in good time to changes.
  • Proactivity: it takes initiative toward goals instead of only reacting.
  • Social ability: it may communicate or cooperate with other agents and people.
  • Persistence: it operates over time, keeping state.
  • Learning or adaptation: it may improve through experience.

Architectures

  1. Reactive architectures: map perceptions directly to actions, as in simple control systems and in behavior-based robots.
  2. Deliberative architectures: keep a model of the world, plan sequences of actions and select the best, a style linked to autonomous planning.
  3. Hybrid architectures: combine fast reactive layers with slower deliberation.
  4. Belief-desire-intention architectures: represent what the agent believes, wants and has committed to doing.
  5. Learning architectures: use reinforcement learning to improve decisions from rewards.
  6. Language model architectures: a model reasons in text, calls tools and keeps memory, with an outer loop that repeats until a goal is met, as covered in agentic AI.

Examples

  • Robots and vehicles: warehouse robots, robotic vacuums, drones, self-driving cars and Mars rovers that navigate, avoid hazards and choose targets with limited contact from operators.
  • Software agents: trading bots, monitoring systems, scheduling assistants and web crawlers.
  • Game agents: programs that learn to play complex games through self-play.
  • Language model agents: systems such as AutoGPT-style programs that break a goal into steps, call tools and loop until finished.
  • Smart infrastructure: autonomous controllers in power grids, data centers and factories.

Levels of autonomy

Different fields define levels. Vehicles are rated from level 0, with no automation, to level 5, full automation in all conditions. For software agents, descriptions run from assistance, through supervised and conditional autonomy, to systems that work for long periods with occasional review. A fully independent system that sets its own ultimate goals is largely theoretical and is the subject of much of the discussion about long-term safety.

Autonomous language model agents

In 2023, projects such as AutoGPT and BabyAGI showed a language model given a goal, creating a task list, executing tasks with tools such as search and code, and repeating. They drew great interest and revealed the weaknesses of early models: they often looped, lost track of the goal, made costly calls or failed on long tasks. Later systems improved planning, memory and reliability, and added limits, approvals and human-in-the-loop checks. See also task decomposition and memory in AI agents.

Benefits

Autonomous agents can work where humans cannot be present or respond quickly, as in space exploration, deep-sea survey and high-speed trading. They scale routine monitoring and operations, work around the clock and free people for judgment-heavy tasks.

Risks and limitations

  • Unpredictable behavior: agents in open environments face situations their designers did not foresee.
  • Goal misspecification: an agent may pursue the stated objective in unintended ways.
  • Compounding errors: small mistakes accumulate over long runs.
  • Safety: physical agents can cause injury, and software agents can cause financial or data harm.
  • Security: attackers may manipulate an agent’s inputs, including through prompt injection.
  • Accountability: responsibility for decisions made without human review can be unclear.
  • Loss of oversight: long runs make monitoring difficult.

Safeguards

Safe designs set clear scopes, limit permissions and spending, test in simulation before deployment, monitor behavior, provide emergency stops, log decisions for audit and keep a human responsible for the overall outcome. Regulators in areas such as road vehicles, aviation and finance impose their own rules.

A worked example

A warehouse deploys fifty autonomous mobile robots. Each senses its surroundings with cameras and lidar, receives pick tasks from a central system and chooses its own path, avoiding people and other robots. When an aisle is blocked, a robot replans without waiting for instructions. If sensors detect a person within a set distance, it stops. Humans set priorities, handle exceptions and can pause the fleet instantly. Productivity rises, and safety incidents are tracked and reviewed weekly.

History

The idea of machines that act for themselves goes back to early cybernetics in the 1940s and 1950s, which studied feedback and control in animals and machines. Robots such as Shakey, built at Stanford in the late 1960s, combined perception, planning and action. The 1980s and 1990s brought behavior-based robotics and the study of software agents and multi-agent systems. Deep reinforcement learning after 2013 produced agents that played video games and board games at expert level, and language model agents emerged after 2022.

Testing and verification

Because autonomous agents face open-ended situations, testing combines several methods: simulation of many scenarios, including rare and dangerous ones; closed-track or sandbox trials; formal verification of safety-critical rules; adversarial testing that tries to confuse sensors or inputs; and staged rollouts with close monitoring. Teams define safe operating conditions, such as speed, area and permissions, and make the agent hand control to a person or stop when it leaves them. Incident analysis after failures feeds back into design.

Ethical and legal questions

When an agent acts on its own, questions arise about who is liable for harm, how decisions can be explained and what level of human oversight society requires. Laws for autonomous vehicles, drones and automated trading already assign responsibility to operators or manufacturers, and new rules for AI systems are extending similar duties. Designers are expected to document decision logic, allow audits and avoid delegating decisions that significantly affect people without meaningful human review.

FAQs

Is every AI agent autonomous? Not fully. Agents vary in how much they decide for themselves and how often they check with people.

Is an autonomous agent the same as a robot? No. Robots are physical, while an autonomous agent may be software only.

Can autonomous agents be trusted? Trust depends on testing, limits, monitoring and the stakes involved.

Key facts

  • Franklin and Graesser defined an autonomous agent in 1996 as a system situated in an environment that senses it and acts on it, over time, in pursuit of its own agenda.1
  • NASA's Mars rovers use onboard autonomy for tasks such as hazard avoidance and selecting science targets, because communication delays prevent real-time control from Earth.2
  • AutoGPT, released in 2023, popularized language model based autonomous agents that decompose goals and loop through actions without step-by-step prompts.3

Related concepts

Quick checkWhat is an autonomous agent?Show answer

It is a system that senses its environment and acts on its own over time to pursue goals, deciding for itself what to do next without continuous human control.

Sources

  1. Franklin and Graesser, Is it an Agent, or just a Program? A Taxonomy for Autonomous Agents, https://link.springer.com/chapter/10.1007/BFb0013570
  2. NASA Jet Propulsion Laboratory, Autonomy for Mars rovers, https://www.jpl.nasa.gov/missions/mars-science-laboratory-curiosity-rover-msl
  3. Significant Gravitas, AutoGPT repository, https://github.com/Significant-Gravitas/AutoGPT

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue