What Is an AI Agent? Agents vs Chatbots Explained

Plain-English definition

An AI agent is a system built around a language model that pursues a goal through a loop of plan → act → observe → continue. Given "find why the build is failing and fix it," an agent might read the error log, search the codebase, edit a file, run the tests, read the new output, and repeat — each step decided by the model, executed through tools.

Four ingredients define the pattern: a model for reasoning, tools for acting, memory of what's happened so far, and a loop that keeps going until done.

Why it matters

Plain chat has a ceiling: the human does all the doing. Agents raise that ceiling: they can complete multi-step work: research that spans dozens of sources, code changes verified by tests, operations tasks that touch several systems. That's why "agentic" capability has become a headline axis on which models are compared, and why benchmarks like SWE-bench (real bug fixing) get so much attention.

How it works

1. Goal and tool definitions go in. The agent receives the task plus a list of tools it may call, each with a description and typed inputs.
2. The model plans and picks an action. Instead of answering in prose, it emits a structured tool call: "run search("build error TS2322")."
3. The system executes the tool, the model never touches anything directly, and appends the result to the running history.
4. The loop continues. With new information in context, the model decides the next step, until it delivers a result or hits a stop condition (success, step limit, budget, or a human checkpoint).

A simple example

"Schedule 30 minutes with the design team next week." A chatbot drafts a nice email asking for availability. An agent calls the calendar tool to read free slots, finds Tuesday 2pm works for everyone, creates the event, and reports back with the invite link. Same model quality — completely different usefulness, because the agent could check and act, not just compose.

Common misunderstandings

• "Agents are smarter models." No, agents are an architecture around a model. The same model can power a chatbot or an agent.
• "Agents are autonomous by default." Good ones are deliberately constrained: scoped tools, permission gates, budgets, and human approval for consequential actions.
• "More steps = more capability." Errors compound. A 95%-reliable step run twenty times succeeds only about a third of the time end-to-end. Reliability engineering is the real work.
• "Agent = one thing." There's a spectrum, from a single tool-calling request to multi-agent networks. Simple usually wins; see the canonical agent patterns.

Where agents appear in real systems

Coding assistants that edit, run, and test code; deep-research tools that browse and synthesize sources; support systems that look up orders and process refunds under policy; and operations copilots that span internal APIs. ModelRefs covers implementation blueprints in the workflow catalog and the hands-on agents tutorial track.

Limitations and risks

• Compounding errors. Long chains multiply per-step failure rates; loops can stall or wander.
• Action risk. A wrong tool call can delete data or send the wrong email. Tool permissions and approval gates are load-bearing, not optional. See guardrails.
• Prompt injection. Agents that read external content can be manipulated by instructions hidden in that content, an open security problem for the whole field.
• Cost unpredictability. Loops consume tokens until stopped; budgets belong in the design.

Next steps

Ready to build? Start with how to build AI agents, then the agents tutorial track and guardrails. Browse proven designs in agent patterns.

Sources and further reading

• Anthropic, Building effective agents — engineering guidance on agent architectures and when to use them.
• Anthropic, AI glossary — provider definitions of agent-related terms.