How to build an AI agent without code

The shortest honest path to a working agent is four nodes and one tool in n8n. According to n8n's AI Agent node documentation, "Connect a chat model and one or more tools, and the agent decides which tools to call to complete a task." The same page adds the one hard requirement: you must connect at least one tool sub-node to an AI Agent node, so an agent with no tool is not an agent in n8n's terms.

Cost is easier to reason about than it looks. The n8n Chat Trigger docs state that one conversation where a user sends 10 messages uses 10 executions, and n8n's pricing page lists the Starter plan at 20€ a month, billed annually, for 2.5K workflow executions. Read that as roughly 2,500 chat messages a month on Starter. It does not include model usage, which is billed separately, but it tells you a personal test agent will not hit the ceiling.

What makes something an AI agent rather than an automation?

An agent decides its own next step; an automation follows the steps you drew. Anthropic's engineering team writes in Building Effective Agents that "Workflows are systems where LLMs and tools are orchestrated through predefined code paths," while agents "dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

In a no-code tool the difference shows up on the canvas. A Zap or an n8n workflow with an AI step is a fixed line of boxes. An n8n AI Agent node is one box with a model, a memory and a set of tools hanging off it, and the model chooses between the tools each time it runs. n8n's docs on what agents do note that when a workflow contains an agent, the agent runs multiple times, for example once to set up, again to call a tool, and again to read the tool's answer and reply. If you are unsure which one your problem needs, the AI agent vs workflow guide walks through the decision, and the AI agent glossary entry covers the vocabulary.

Which no-code tool should a beginner start with?

Start with n8n if the goal is to learn, because every part of the agent is a separate node you can inspect. The free and entry options below were checked on each vendor's own pricing or docs page on 12 October 2026.

Tool	Free way to start	First paid tier	What a beginner builds
n8n	14-day Cloud trial, 1,000 executions, no credit card; self-hosted Community Edition on GitHub	Starter: 20€ a month billed annually, 2.5K executions	AI Agent node with a chat model, memory and tools as visible sub-nodes
Zapier Agents	Agents Free: 400 activities per month	Agents Pro: $33.33 a month billed annually ($400 a year), 1,500 activities	An agent described in plain language that acts in the apps you connect
Make	Free plan: 1,000 credits per month	Not compared here; see Make's pricing page	Make AI Agents (beta), listed on every plan

The n8n trial terms come from n8n's trial documentation, which gives 14 days of Pro features with "a limit of 1000 executions." The Zapier figures come from the Agents plans section of Zapier's pricing page, where an activity counts actions the agent takes, web browsing, and lookups in attached knowledge. Make's pricing page lists Make AI Agents (beta) with the note that you can use Make's AI provider on all plans or your own LLM key.

Zapier is the faster setup if your work already lives in apps it connects. Its pricing FAQ is clear about the boundary: "Zapier Agents can only take actions in the apps you've connected and use the triggers and actions you've set up." n8n asks a little more of you up front and gives back a clearer picture of why the agent did what it did. For a side-by-side of the two older automation products, see Zapier vs Make for product managers; if you have never opened n8n, n8n workflows for product managers covers the editor and a first workflow with an AI step.

How do you build your first agent in n8n, step by step?

Build it in this order, and test after each step rather than at the end.

1. Add a Chat Trigger. Create a new workflow and add the Chat Trigger node. Leave "Make Chat Publicly Available" off while you build; n8n's docs say to turn it on only when you are ready to publish.
2. Add the AI Agent node and connect it to the trigger. Since n8n 1.82.0 every AI Agent node works as a Tools Agent, so there is no agent type to choose.
3. Attach a chat model. Click the Chat Model connector and pick a model. On n8n Cloud Starter or Pro, Gateway credits let you use models from OpenAI, Anthropic or Google Gemini without a provider account or API key. On self-hosted n8n, add a credential for your own provider instead.
4. Attach Simple Memory. Click the Memory connector and choose Simple Memory. Set Context Window Length, the number of previous interactions the agent keeps. Keep it small for a test agent and raise it only if answers lose the thread.
5. Attach one read-only tool. For a first test, use the Wikipedia tool, which lets the agent search and return information from Wikipedia; its n8n docs page lists no credential to set up. One tool makes it obvious whether the agent is choosing to call it.
6. Write a System Message. Open the AI Agent node's options and add a System Message, which n8n describes as a message sent to the agent before the conversation starts, used to guide its decisions.
7. Test in the chat panel. Open the chat, ask one question the tool can answer and one it cannot, and read the execution log to see whether the tool was called.

Step 7 is the one most beginners skip. An agent that answers well without calling its tool has not proven anything yet, because the model may simply be answering from its training data.

What should the first system message say?

The system message should name the job, the one tool, and what to do when the tool has no answer. Here is a starting point for a product team's glossary helper:

You help a product team understand unfamiliar technical terms.
For any term you are asked about, search Wikipedia first and base your answer on what you find.
Answer in three sentences or fewer, in plain language for a non-engineer.
If Wikipedia has nothing relevant, say so. Do not guess.

Keep it short and specific. The AI Agent node's Tools Agent options also include Max Iterations, which defaults to 10 and caps how many times the model runs to produce an answer, and Return Intermediate Steps, which adds the agent's tool calls to the output so you can see its reasoning trail. Turn on Return Intermediate Steps while testing.

How do you connect the agent to your real data safely?

Swap the test tool for a read-only connection to the data you actually care about, then add write access last. n8n's docs on how tools work list the Call n8n Workflow Tool, the HTTP Request Tool and a registry of MCP servers as ways to give an agent access to other systems, and include a "Chat with Google Sheets" example built on the Call n8n Workflow Tool. A spreadsheet of feature requests or a support FAQ is a good first real source, because reading it cannot break anything.

When a tool can send, edit or delete, put a person in front of it. n8n's human-in-the-loop documentation describes the mechanism: when the agent wants to use a tool that has human review enabled, "The workflow pauses and sends an approval request through your configured channel," and a reviewer approves or denies the exact call. The same page suggests starting with human review on, then reducing oversight as confidence grows. The tool use glossary entry and the human-in-the-loop entry explain both ideas in more depth.

What does Simple Memory actually remember?

Simple Memory keeps a window of recent chat history for the current session and nothing more. The Simple Memory node docs give it two settings: a Session Key, which decides which conversation the history belongs to, and Context Window Length. It does not learn facts across users or remember last week.

One constraint matters before you share the agent. n8n's docs warn that if your instance uses queue mode, Simple Memory does not work in an active production workflow, because n8n cannot guarantee every call reaches the same worker. n8n's docs describe queue mode as a setting for scaling self-hosted instances, so it rarely affects a first test agent; for a team rollout, move to one of the database-backed memory nodes. The agent memory glossary entry covers the difference between session memory and long-term memory.

When is a no-code agent the wrong thing to build?

Often, and the honest objection is that most first agent ideas are workflows in disguise. Anthropic's guidance in Building Effective Agents is blunt: "we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all." If the steps are the same every time (new form entry, summarise it, post it to Slack), a plain automation is cheaper and more predictable than an agent.

Non-determinism is the second cost. Zapier's own Agents FAQ answers "Why don't I get the same outcome from AI every time?" by pointing to the non-determinism of language models. An agent that gives a slightly different answer each run is fine for research and drafting, and a problem for anything a customer or a finance system depends on. Use an agent where the path genuinely varies by input, such as triaging mixed questions or researching across sources, and the which tasks to automate guide for everything else.

What should you build after the first agent works?

Add a second tool only after the first one has a question it reliably answers. Write down five test questions, two that need the tool, two that do not, and one the agent should refuse, and rerun them every time you change the system message or the model. The guide to writing evals for AI products turns that habit into a repeatable test set, and how to design an AI agent covers scoping tools, guardrails and when an agent needs more than one.

If you want structured practice on the design side, the AI Agents bootcamp is a 2-week programme with 3 live sessions and 9 self-paced microlessons on agent fundamentals, tool use, memory, human-in-the-loop design and evaluation, led by its own programme director and also available self-paced.