Artificial Intelligence Agents

An autonomous AI agent is software that observes an environment, chooses actions to pursue an objective and executes them with limited human intervention. Unlike a model that only returns an output for a prompt, an agent operates through a continuing perception-action loop, using the results of earlier actions to guide later ones.

How an agent operates

An agent repeatedly converts information from its environment into decisions and actions.

• It perceives inputs such as user instructions, files, databases, application data, system messages and results returned by digital tools.
• It may combine an AI model, programmed rules and stored context to select actions or create and revise a plan.
• It acts through software interfaces, including application programming interfaces, by retrieving information, generating content, updating records or initiating workflows.
• It observes the result and continues the feedback loop until it reaches its goal, a stopping condition or an intervention threshold.

Difference from rule-based software

Conventional rule-based software typically applies explicitly programmed conditions to specified inputs and follows predefined workflows. An autonomous agent instead selects and sequences actions in pursuit of a goal, using environmental feedback to determine what to do next.

• Agents may handle changing situations by comparing alternatives, revising plans and maintaining memory or state across steps.
• Their behaviour may be less predictable because model-based decisions can vary with context, data and earlier actions.
• The distinction is not absolute: an agent may contain programmed rules, while its autonomy remains bounded by its objective, tools, permissions, operating rules and computational resources.

Control and oversight

Sequential action creates special risks because one error may influence later decisions or propagate across connected systems.

• Human oversight may include approval requirements, escalation rules and the ability to interrupt or override actions.
• Least-privilege permissions and restricted tool access limit what the agent can read, modify or execute.
• Logging and monitoring make actions traceable and help identify abnormal behaviour, repeated failures and unintended consequences.
• Testing should assess behaviour across foreseeable operating conditions, not merely the accuracy of the underlying model.