What is an AI Agent?

Beyond a single prompt

An AI agent is a system that uses an LLM (or other models) to pursue a goal over multiple steps—not just answer one question. It can plan, use tools, remember context, ask for clarification, and adapt when something fails.

Think of the difference between “summarize this PDF” (one shot) and “research competitors, update the CRM, and draft a briefing” (ongoing, tool-mediated work). The latter is agent-shaped.

Typical components

Agents are architectures, not a single model call. Common pieces include:

• Policy / planner — decides what to do next (often the LLM with a structured prompt).
• Tools — functions the agent can invoke: search, APIs, code execution, databases, MCP servers.
• Memory — short-term (conversation) and long-term (vector stores, user profiles, run history).
• Environment — the systems and data the agent can observe or change.
• Guardrails — limits on actions, approvals, and output validation.

Single agent vs multi-agent

A single agent with good tools can handle many workflows. Multi-agent setups assign roles—researcher, coder, reviewer—and pass work between specialized agents. That can improve quality but adds coordination cost, latency, and debugging complexity.

Start with the simplest design that meets the task; add agents when clear separation of concerns justifies the overhead.

Agents vs chatbots

Chatbots optimize for conversational UX. Agents optimize for task completion. Overlap exists—many products are conversational agents—but the engineering focus shifts toward reliability, tool contracts, observability, and recovery from errors.

When agents make sense

Agents are valuable when work is open-ended, requires external systems, or benefits from iteration. They are a poor fit when a deterministic script or classical ML model is cheaper, faster, and easier to audit.

The harness—the runtime that hosts the agent—is what turns a clever demo into something you can ship. That’s covered in the next article.