What LLM agents are, how they differ from plain LLMs, the core components (tools, memory, planning), and how to build one that works in production.
An llm agent is a language model in a loop: it calls tools, reads the results, decides the next move, repeats until the goal is hit. A plain LLM answers once and stops. An agent works the problem.
That's the whole difference. Answering vs acting. Everything else, frameworks, memory, planning, is plumbing around that loop.
Plain LLM call: text in, text out. Pure function. No memory, no side effects. Good for drafting, classifying, extracting.
Agent adds four things:
Deciding which you need? We broke down the honest version in autonomous agents: what actually works vs hype. Fixed sequence of steps = workflow, not agent.
model. the reasoning engine. tool-calling reliability beats benchmark scores. small local models work for narrow agents, degrade fast as tool count grows.
tools. typed functions with descriptions the model reads: get_invoice(id), search_crm(query). the model picks tools by reading descriptions. vague descriptions = vague agent. MCP made this way easier, examples in mcp-powered agents in n8n.
memory. context window is short-term memory and it runs out mid-loop. production agents need a keep/summarise/externalise strategy. five patterns that survive real use: agent memory patterns.
the loop. reason, act, observe, repeat. ReAct-style. skip the elaborate planning frameworks, they demo well and fail weird. hard limits mandatory: max steps, max cost, max time. every runaway agent bill = missing stop condition.
guardrails. anything touching customers or money gets an approval gate. agent researches and drafts, human approves, workflow executes. 90% of the time savings, none of the incidents.
Production agent we run for invoice triage. Goal: match supplier invoices to purchase orders, flag mismatches.
search_purchase_orders, finds candidatesAgent is justified here because invoices are messy: partial deliveries, price changes, bundled lines. A fixed workflow needs hundreds of branches. The agent reasons per case and shows its work.
Most expensive mistake of 2026: building an agent where a workflow would do. Known steps = workflow. One transformation = single LLM call. Agent only when the path genuinely varies per case.
And one agent before several. Multi-agent systems fail in ways that are brutal to debug, patterns that hold up are in multi-agent orchestration.
No framework needed. n8n agent node + 2-3 well-described tools = working agent in an afternoon. Full setup with approvals and error handling: ai agent workflows in n8n. Start narrow, log every step, widen scope only after a week of real inputs.
Need one built? we ship production llm agents for clients →