What are AI Agents?

Definition

An AI agent is a software system that uses artificial intelligence to autonomously pursue goals by perceiving its environment, reasoning about actions, executing those actions, and evaluating results. Unlike simple chatbots that respond to single prompts, agents maintain context across interactions and can take multiple steps to accomplish complex tasks.

The key distinguishing characteristics of agents are: - Goal-directed behavior: Agents work toward specified objectives, not just respond to inputs - Autonomy: Agents can operate with minimal human intervention (degree varies) - Environmental interaction: Agents perceive and modify their environment through tools - Iterative refinement: Agents evaluate outcomes and adjust their approach

Agents vs. Chatbots vs. Scripts

Understanding what agents are requires distinguishing them from related systems:

Scripts are deterministic programs that follow fixed instructions. Given the same input, they produce the same output. They cannot adapt to unexpected situations.

Chatbots use language models to generate responses to user inputs. They may be stateless (each message is independent) or maintain conversation context, but they respond rather than pursue goals.

Agents combine language model capabilities with goal pursuit, environmental interaction, and iterative refinement. They decide what actions to take, execute those actions, and adjust based on outcomes.

The boundary between these categories is not always sharp. A sophisticated chatbot with function calling capabilities begins to exhibit agent-like properties. The key question is whether the system autonomously pursues goals across multiple steps.

The Agentic Loop

Agents operate through a continuous cycle called the agentic loop:

1. Perceive — The agent gathers information about its environment, task state, and available resources. This might include reading documents, querying APIs, or parsing user instructions.

2. Reason — Using this information, the agent reasons about what action to take next. This involves understanding the current state relative to the goal, evaluating options, and selecting an approach.

3. Act — The agent executes the chosen action. This could be calling a tool, generating content, making an API request, or delegating to another agent.

4. Evaluate — After acting, the agent assesses the result. Did the action advance the goal? Were there errors? Should the approach change?

This loop repeats until the goal is achieved, a failure condition is met, or human intervention is requested. The sophistication of each phase varies by agent design—some agents have elaborate planning, others rely on simple heuristics.

The Autonomy Spectrum

Agents exist on a spectrum from highly supervised to fully autonomous:

Human-in-the-Loop — The agent proposes actions but waits for human approval before execution. Maximum safety, minimum speed. Appropriate for high-stakes decisions.

Human-on-the-Loop — The agent executes autonomously but humans monitor progress and can intervene. Balances efficiency with oversight.

Human-out-of-the-Loop — The agent operates independently, with humans only involved in initial configuration and final review. Maximum efficiency, requires high confidence in agent reliability.

Most production systems use hybrid approaches: autonomous execution for routine operations with human checkpoints for critical decisions. The appropriate level of autonomy depends on task risk, agent reliability, and consequence severity.

Core Agent Capabilities

Effective agents typically combine several capabilities:

Planning — Decomposing high-level goals into actionable sub-tasks. Good planning improves efficiency and reduces errors.

Memory — Maintaining context across interactions. Short-term memory holds current task state; long-term memory stores learned information.

Tool Use — Interacting with external systems through defined interfaces. Tools extend agent capabilities beyond text generation.

Orchestration — Coordinating multiple agents or steps in complex workflows. Orchestration handles dependencies, failures, and resource allocation.

These capabilities are covered in detail in the Core Terms section of this reference.

What This Reference Covers

This Agents Reference provides:

Core Terms — Foundational definitions for agent concepts including agentic workflow, the agentic loop, tool use, memory, planning, orchestration, and human-in-the-loop patterns.

Architecture Patterns — Proven approaches for structuring agent systems, including planner-executor, router-supervisor, and reflect-revise patterns.

Esy Workflows — Practical implementation examples showing how these concepts apply in Esy's research automation platform.

Each entry includes why the concept matters, common misconceptions, and cross-references to related topics.