How to build an AI agent without coding: step by step

The short version: building an AI agent takes neither code nor a server, but five decisions — who it is, what its task is, how it works, what the answer should look like and which tools it needs. Those decisions go into its instructions (the system prompt), the tools are ticked in a list, and ten minutes later the agent is working. Below: each step with examples, and a builder that puts the agent together for you.

Step 0. Pick a task the agent will definitely manage

The most common reason for disappointment is not a bad platform but a poor first task. A good agent task repeats, has several steps and has a result you can check.

On the left are tasks where it is clear what the agent must do and how to check the result. On the right are goals, not tasks: there is nothing for the agent to get hold of. If you are short of ideas, look at the AI agent examples — there are nine ready ones.

Step 1. Role and task

The first two lines of the instructions set who the agent thinks it is and why it exists. The role narrows its style and knowledge ("an employment lawyer" answers differently from "an assistant"); the task sets the boundaries.

Step 2. How it works

The agent chooses its own steps, but a suggested order makes it far more consistent: it will not skip the check or start writing the report before it has the data. Three to five steps are enough — this is a route, not a rulebook.

How you work:
1. Search for 5–7 recent sources.
2. Open the most useful ones and read them in full.
3. Combine the data and note where sources disagree.
4. If figures differ by more than 10%, give both values.

Step 3. Answer format and rules

The format is what you get at the end: a table, a post, a document, a short answer. The rules are what the agent must not do. Rules are what protect you from a model's main weakness — confident invention.

• "Never invent figures, prices or dates. If the data is missing, say so."
• "Every fact from the web comes with a link to its source."
• "If information is missing, ask a clarifying question first."
• "Do not promise discounts or deadlines — the account manager decides those."

Step 4. Tools: exactly as many as needed

Tools turn a model into an agent: without them it can only talk. But every extra tool is one more description in the context and one more chance to pick the wrong one. The rule is simple: start with one to three and add more when the agent turns out to be missing something.

Build an agent right now

The builder below assembles instructions from the five blocks and shows straight away what is missing. Start from a template or from scratch. The "Build the agent in Twelver" button opens the editor with the name, prompt, description and tools already filled in — all that is left is to press Save. You can also just copy the finished prompt: it works on any platform.

You are a market analyst.

Your task is to find up-to-date data on the user's topic and compare the options.

How you work:
1. Search for 5–7 sources.
2. Open the most useful ones and read them in full.
3. Combine the data and note where the sources disagree.

Answer format: a table followed by three short takeaways.

Tools and when to use them:
- Search the web: when you need fresh prices, news or competitor data.
- Write documents and tables: when the output is a report, an article, a table or a brief.

Rules:
- Every fact from the web comes with a link to its source.
- Never invent figures, prices or dates. If the data is missing, say so.
- Answer in English, even when the sources are in another language.

Step 5. Knowledge, memory and scheduling

The basic agent is ready. Three add-ons make it genuinely useful:

• Knowledge base. Attach a price list, a policy or a handbook to a chat with the agent — the files are kept in its knowledge base, and the agent searches them every time it runs. That is the answer to "how do I train an AI agent on my own data": you do not retrain the model, you give it documents.
• Memory. The agent writes down important facts itself — your name, tone, what you agreed — and remembers them in new chats. How that works is shown in the chapter on the personal AI assistant.
• Scheduled runs. The agent starts itself at the right time and sends the result as a notification. A schedule builder is in the agent examples.

Step 6. Test the agent with three requests

Before you trust an agent with real work, run three tests. They take five minutes and catch most problems with the instructions.

If a test fails, fix the instructions, not the wording of your request: the agent will be used by people who do not know your tricks.

How to build an AI agent in Twelver, step by step

1. Open a chat and, in the agent picker, choose "Create New Agent".
2. Give it a name, an avatar and a short description.
3. Paste the instructions into the "System Prompt" field.
4. In the "Capabilities" section, open a category and tick the tools.
5. Save the agent and send it its first task.
6. Add files to a chat with the agent — they will appear in its "Knowledge base" section.

You can share an agent you have built by link: the link shows its name, icon and description, while your instructions and knowledge-base files stay private.

How to build an agent on other platforms

The five blocks of the instructions are universal — only the place you paste them changes.

• n8n. The AI Agent node: the instructions go into the System Message, the model is connected as a separate node, and the tools as Tool nodes. Best when the agent has to react to events in other services.
• Botpress. A visual builder: the instructions go into the agent's settings, and its knowledge comes from uploaded files or pages. Best for a support agent on a website.
• ChatGPT. Custom GPTs with instructions and files; agent mode is on paid OpenAI plans.

A detailed comparison, with a picker that matches a platform to your task, is in the ranking of AI agent platforms, and the options that cost nothing are in the breakdown of free AI agents. For the basics, go back to the overview: what an AI agent is.

Five mistakes people make with their first agent

1. The role is "assistant". Too general a role gives too general answers. Name the profession and the specialism.
2. Every tool at once. The agent starts drawing a picture where you asked for a table. Keep the ones you need.
3. No rule about sources. An agent that searches the web without being required to give links is an agent you cannot check.
4. Three screens of instructions. Models follow long, contradictory instructions worse than short ones. If the agent has to do too much, split it into a team of agents.
5. No tests. The first real request should not be the first request, full stop.