Beginner friendlyNo code path includedDeveloper path included

Create your first agent

This guide helps you build one useful agent, test it with realistic messages, and decide how to deploy it. You can use the dashboard only, or use HiveLang if you want code-level control.

Step 1: choose one job

A good first agent should not do everything. Give it one job, one audience, and clear boundaries.

Too vague

Be a helpful AI assistant.

Better

Help customers understand failed payments, ask for missing details, and guide them to retry safely.

Step 2: build without code

  1. 1

    Open the builder

    Go to Dashboard - Bots - New bot. Choose HiveMind for a guided prompt flow, or Builder for manual setup.

  2. 2

    Describe the job clearly

    Use plain English. Include who the agent helps, what it should do, what tools it can use, and what it must not do.

  3. 3

    Connect only needed tools

    If the agent does not need Gmail, Calendar, GitHub, billing, or a custom API, leave it disconnected. Fewer tools usually means faster and safer agents.

  4. 4

    Test in the playground

    Try a normal request, a vague request, a tool-based request, and a request the agent should refuse.

Step 3: optional HiveLang version

Developers can define the same behavior in HiveLang. This makes the agent easier to review, version, and deploy from a repo.

bot SupportGuide {
  description: "Answers common product questions"

  instructions {
    You are a clear support agent.
    Ask for missing context before guessing.
    If you cannot verify something, say what you need next.
  }

  on user.message {
    respond
  }
}

What this code means

  • bot SupportGuide names the agent.
  • description explains what the agent is for.
  • instructions define behavior and safety boundaries.
  • on user.message handles incoming user messages.
  • respond tells the runtime to generate the answer.

Step 4: deploy one surface

Start with one surface. Add more only after the first one works.

Hosted web link

Share a public agent page quickly.

Telegram

Let users message your agent in Telegram.

API

Call the agent from your backend.

Step 5: verify with observability

After a test or deployed run, open the trace and check:

  • what the user asked
  • what context was loaded
  • which model ran
  • which tool was selected
  • what arguments were used
  • how long it took
  • what to fix if it failed

Next reads

How helpful was this page?

Not helpfulExcellent