BETA
An AI agent is a system in which a model plans and executes multi-step tasks by calling tools, reading results, and iterating toward a goal.
The core of an agent is a loop: the model receives a goal, decides on an action — search, run code, edit a file, call an API — observes the result, and repeats until the task is done. Coding agents, deep-research assistants, and computer-use agents all instantiate this pattern, and by 2025 it had moved from demos into production, with coding agents the clearest commercial success.
Reliability rests on more than the model: tool design, context management across long sessions, checkpoints, and permission boundaries determine whether an agent completes a fifty-step task or wanders off after step ten. Standards like the Model Context Protocol emerged to make tool integration portable across models and applications.
Agents are the most token-hungry workload category by far. A single task can span dozens of model calls with an ever-growing context resent each step, so agent costs run one to three orders of magnitude above single-prompt usage. Prompt caching, context pruning, and routing sub-steps to cheaper models are what keep agent economics viable.
Last revised 2026-07-05 · All glossary terms → · Live AI model pricing →