“An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators.”Stuart Russell, Peter Norvig · Artificial Intelligence: A Modern Approach · 1995 · Artificial Intelligence: A Modern Approach, ed. a 4-a (2020), cap. 2 «Intelligent Agents», §2.1 — definiția agentului
An agent is a perceive–decide–act loop, and the guardrails sit on every turn.
The textbook definition maps exactly onto what runs today: the sensors are the context (the user's message, files, tool results), the actuators are tool calls (search, run code, send an email). The loop closes when a tool result comes back into the context and the model chooses the next step. That is the difference from a chat call: a call never sees the consequence of its own action. Autonomy has a plain measure — how many turns the loop makes without a human. An agent is defined by its tools and its stopping rule, not by its model.
The card says where to put the fences. The next step is different: who stops the loop when nobody is around? A stopping rule is not a safety patch bolted on at the end; it is part of what defines an agent. An agent with no clear termination criterion is not autonomous, it is merely unrolling. Think of a thermostat: it asks no one when to stop heating. It has a condition, the target temperature, and stops on its own. Likewise, an agent that sends emails needs a verifiable condition: the message was sent, the confirmation arrived, the loop closes. Without that, every turn adds fresh risk, and errors compound. The question now opening: what happens when the stopping condition is ambiguous, and the agent must judge for itself whether it is done?
When the stopping condition is ambiguous, the agent must judge, and judgment happens over perceptions. Here lies a trap: a tool's result is not the fact, only a report of a fact. Code that runs without errors does not mean the program is correct. A search returning a page does not mean the page is true. An agent that treats perceptions as reality closes its loop on an illusion, and later turns build on it. Think of a dishwasher checking a plate by looking at only one side. His loop closes too early. A good agent always asks: could my perception be misleading? But this raises another problem: verification costs time and actions, so who decides how much checking one turn deserves?
Verification costs something, so the agent must weigh it. The textbook calls this rationality: choosing the action that maximizes a performance measure, not the perfect action. An agent that verifies everything forever is not safer, it is paralyzed. One that verifies nothing is reckless. Between them lies the judgment of how much one turn is worth. Think of someone buying a used car: they don't hire a mechanic for every bolt, but they don't buy without a test drive either. They pick the two or three checks that buy the most confidence. For agents, this means deciding upfront which turns are cheap and reversible, and which are costly and hard to undo. Verification belongs before irreversible steps, not after every step. But what if a step looks reversible and isn't? That is the question this opens.
Why it matters Whoever sees the agent as a loop knows where the guardrails go: on every turn, not just at the entrance.