Definition
An AI agent is a system that lets a model influence the sequence of actions used to pursue an objective. It can inspect context, choose a tool, receive an observation, revise a plan and continue until a stopping condition is met.
Difference from a workflow
A workflow follows paths substantially defined in advance. An agent chooses among possible paths at runtime. Real systems often combine both: deterministic code controls permissions and irreversible actions while a model handles interpretation, planning or recovery.
System boundary
The agent includes more than the model. Tool definitions, credentials, memory, environment, logs, budgets, approval gates and evaluators all contribute to its behavior.
Electronic Artefacts position
Electronic Artefacts connects agents to contextual execution. An action should be constrained by actor identity, target entity, relation, evidence level, visibility and current state. VASTE provides a relevant research context for that model.
Limitations
Longer action loops increase latency, cost and the chance of compounding errors. Agents need sandboxing, bounded authority, explicit completion criteria and human review where consequences are difficult to reverse.
References
See Building Effective Agents, Contextual Execution, Large Language Model and Autonomous System.