4. What Is an Agent?

Definition. The word agent is old in AI. In classical AI, an agent is any system that senses its environment and acts on it to reach a goal (1). That covers everything from planning programs to robot vacuums. This course focuses on one type: the LLM-based agent. Here, a large language model decides what to do next, instead of hand-coded rules or a separately trained policy. The rest of this manual covers the runtime that turns a single model call into such an agent.

A single LLM call is not an agent.

User input
    |
    v
+-------------+
| LLM Engine  |
+-------------+
    |
    v
Text output

A single call transforms context into text. It may be useful, but it does not repeatedly act, observe, update state, or pursue a multi-step goal.

An agent is a system that repeatedly uses an LLM to decide what to do next. It acts through tools, sees the results, updates its state, and keeps going until a stopping condition is met (1).

Goal
  |
  v
+------------------------+
|      Agent Harness     |
|                        |
|  Context + Loop + Tools|
|                        |
+-----------+------------+
            |
            v
     LLM Engine call
            |
            v
     Proposed action
            |
            v
     Tool execution
            |
            v
     Observation
            |
            v
     Update context/state
            |
            v
     Continue or stop

A minimal agent requires five elements:

1. Goal, the task the agent is trying to complete.
2. Observation, the information the agent has at each step (context, tool results, environment state).
3. Reasoning, analysis and decision making, including deciding whether the goal has been reached (the stop condition).
4. Action, at least one way to affect the world or fetch information.
5. Memory, state that is kept across steps (short-term context and long-term storage).

If reasoning has no stop condition, the system is not a production-ready agent. It is a loop that can run forever.

Autonomy Spectrum

Agents exist on a spectrum (2)(3):

Mode	Human role	Typical use
Suggested action	Human reviews every action	High-risk decisions
Approval gates	Human approves dangerous actions	Mutations, payments, deletions
Bounded autonomy	Agent acts within strict limits	Search, summarization, constrained operations
High autonomy	Agent runs with monitoring	Sandboxed analysis, repetitive workflows

More autonomy requires stronger guardrails, better observability, and clearer rollback strategies.