Relevance AI: How to Build Your First AI Agent

Most people open Relevance AI, see a blank agent, and stall. The platform is capable, and that is exactly the problem on day one: there are several right ways to start and no obvious first move. This guide picks one clean path and walks it end to end.

The single biggest predictor of whether your first agent works is scope. Pick one job with a clear input and a clear output, like read inbound leads and write a short qualification note. Resist the urge to build the agent that runs your whole business. You can add responsibility later once you trust the first version.

What do you need before you start?

You need an account, a decision about what the agent will do, and access to whatever it needs to touch. Relevance AI offers a free tier alongside paid plans, so you can build and test without committing budget on day one.

Write your job description before you open the builder. Two or three sentences is enough: what triggers the work, what the agent should produce, and what it must never do on its own. That short brief becomes the backbone of your prompt later, and it stops you from redesigning the agent halfway through.

Which of the three build paths should you pick?

Relevance AI's documentation describes three ways to create an agent, and they suit different starting points. Pick by how well defined your task already is.

• Invent: you describe the task in plain language and the platform generates a working agent setup for you to review and refine. Best when you know the outcome but not the mechanics.
• Clone from the Marketplace: you start from a pre-built agent someone has already tested, then customise it and connect your own integrations. Best when your task is a common one.
• Build from scratch: you write the prompt, add tools, and optionally attach knowledge yourself. Best when your workflow is specific and you want full control.

A practical middle route is to use Invent to produce a first draft, then edit it as if it were your own build. You get a sensible structure in seconds and you still learn where every setting lives, which matters the first time something misbehaves.

Build your first agent from scratch

1. Open Agents and create a new one — Go to the Agents area of the platform and start a new agent. Give it a name that says what it does, not a cute one. Relevance AI's own guidance is that a purpose-led name works better than a generic bot name.
2. Write the agent summary — A short description of what the agent does and what it is expert in. This is not decoration, it shapes behaviour and sets expectations for anyone who later opens the agent.
3. Choose the language model — You can let the platform pick for you, or select manually. Their documentation lists models from OpenAI, Anthropic, Azure OpenAI, Google Gemini, and OpenRouter. Start with the automatic option unless you have a reason not to.
4. Write the prompt — The standing instructions the agent follows on every task. Cover responsibilities, rules, tone, and what to do when it is unsure. Reference tools by name inside the prompt so the agent knows what it is allowed to reach for.
5. Add the tools it needs — In the Tools section of the agent builder you can add pre-built tools or your own. Add only what this job requires. Every extra tool is one more thing the agent can get wrong.
6. Attach knowledge if the job needs facts — If the agent must answer from your documents or data, add knowledge and choose whether it goes into the prompt or the agent searches it. Skip this entirely if the task does not need it.
7. Test it in chat before automating — Talk to the agent, give it three or four realistic inputs, and read what it actually does. Fix the prompt until the behaviour is boring and predictable.
8. Add a trigger last — Only once it behaves in chat should you connect a trigger so it runs on its own. Triggers are configured per agent and can be paused without deleting them.

How do you write a prompt that actually holds?

Write the prompt the way you would brief a new hire on their first morning. Relevance AI's documentation frames it the same way: the prompt is the set of instructions the agent follows every time it works on a task, covering responsibilities, expectations, rules, and communication style.

Be specific about the boundaries, not just the goal. A prompt that says "qualify inbound leads" leaves the agent guessing. A prompt that says "score each lead 1 to 5 against these three criteria, never contact the lead, and flag anything scoring 4 or above for a human" produces the same answer twice in a row.

Their prompt editor supports a slash menu for inserting tool references, placeholders, and formatting, and you can view the prompt as Markdown or raw text. Naming your tools clearly pays off here, because tool names end up in front of the model and vague names lead to the wrong tool being chosen.

How do you give the agent tools without losing control?

Tools are what turn a chatbot into something that does work. In the agent builder you add tools from the pre-built set or from tools you have created, then configure how each one behaves at run time.

The setting that matters most on a first build is the approval mode. Their documentation describes three: Auto Run, where the tool fires without interruption, Approval Required, where the agent explains its intent and waits for you, and an option that lets the agent decide when approval is warranted.

There is also a failure policy to set. Relevance AI's tool settings let you retry errored tools, cap the retry count, and choose whether a failed task ends or escalates to a human. Escalation is usually the better default for anything that touches a customer.

If assembling the prompt, the tools, and the approval rules feels like a lot of setup before any work happens, that is a fair reaction, and it is the part we removed at Sistava. You hire an AI Employee for the role, tell it in plain English what you want handled, and it arrives with the tools already connected and asks you when it is unsure. There is no builder step between you and the first result.

How do you make the agent run without you?

Triggers are the answer, and they are configured per agent. Their documentation lists integration triggers across a very large catalogue of services, scheduled triggers for time based runs, custom webhooks, direct API triggers, the JavaScript SDK, and tools used as triggers.

You reach them from the agent's own settings, where you add a trigger and configure its conditions. An agent can carry more than one trigger, each running independently, and you can pause and resume them without losing the configuration. That pause control is worth knowing on day one, because it is how you stop a misbehaving agent without dismantling it.

Add the trigger only after the agent behaves correctly in chat. An agent that is fifty percent right in a conversation is one hundred percent wrong when it fires unattended twenty times overnight.

What does running an agent cost?

Relevance AI's pricing documentation, checked on 17 August 2026, describes plan tiers named Free, Pro, Team, and Enterprise, and two separate usage currencies. Actions are counted when an agent executes a tool, and vendor credits cover the underlying model and tool usage.

The practical consequence is that a chatty agent with many tool calls costs more than a terse one, so tightening your prompt is a cost lever as well as a quality lever. Their admin documentation also covers usage alerts and spend controls, which are worth turning on before your first automated trigger goes live. Check their current pricing page for the numbers, since they change.

What goes wrong on a first build?

• Scope creep in the prompt: one agent asked to do five jobs does all five poorly. Split it, or move to a multi-agent setup later.
• Vague tool names: the model chooses tools partly by name, so "Tool 2" is a trap. Name tools after the action they perform.
• Automating before testing: a trigger attached to an untested agent turns a small prompt error into a large cleanup.
• No stop condition: tell the agent explicitly what to do when it does not have enough information, otherwise it will invent a path forward.

None of this is a knock on the platform. Relevance AI gives you real control, and real control means real decisions. If you would rather skip the decisions and get the outcome, that is the trade Sistava makes: you hire an AI Employee, brief it in a paragraph, and it sets itself up. Different tool, different philosophy, and the right choice depends on whether you enjoy building.

Sources

• Relevance AI documentation, Introduction
• Relevance AI documentation, Create an Agent
• Relevance AI documentation, Prompt
• Relevance AI documentation, Tools for agents
• Relevance AI documentation, Triggers
• Relevance AI documentation, Pricing
• God of Prompt, How to create an AI agent with Relevance AI
• BenAI, Build your first no-code AI agents