KilatBot Articles

Articles

How to create your own AI agent without code

· · Updated

To create an AI agent without code in KilatBot, start from a template or a blank agent, then shape its instructions, skills, tools, and schedules around one job. You are configuring an assistant for your business, not training a new model or building an agent platform.

Where do you start when creating an AI agent?

Write one sentence describing what you want back. “Help with operations” is too broad to check. “Turn these work notes into a draft handover, with open questions separated” gives you something to inspect. Choose work you understand well enough to notice mistakes.

Prepare a small input and an example of an acceptable answer. Mark which details must match exactly and which parts may be rewritten. Decide what the agent should do when a detail is absent. Ask for a question or a clearly marked gap rather than a guess.

This preparation matters more than finding an impressive name for the agent. If you cannot describe a useful result, adding more skills will not tell you whether the task worked.

What does a template set up?

A KilatBot template sets the persona, skill, and schedules. You can change them later or start blank. Choose the starting point closest to your job, then review what it includes. Follow the Quickstart for the exact creation steps. To see everything an agent can do, read the agent overview.

Give the agent a name and choose its initial language. Read the review before creating it, especially its skills, schedules, and channels. Do not assume a template's name means every task in that field is covered.

A template's schedules start paused. Leave them paused while you test the job. The agent starts on Kilat AI, so a separate model key is not needed. If you later add your own provider key, that provider bills the usage.

What rules should you write for the agent?

Use the persona to describe how the agent should work for you. Include the source it should trust, the answer format, and when it should ask a question. Keep important rules concrete enough that you can check an answer against them.

Example: “Help me prepare internal handovers. Use the notes I provide. Separate completed work, unfinished tasks, and questions. Keep names and dates unchanged. Mark missing owners. Produce a draft for my review, not a message to send.”

Test it with a complete note and an incomplete one. If the agent invents an owner or deadline, correct the instruction and try that case again before relying on it.

Avoid conflicting rules such as “always finish the report” and “never proceed without missing figures”. Decide which matters more. For business records, prefer a visible gap over a complete-looking answer you cannot support.

Separate skills from permission.

Review the skills supplied by the template and keep the setup focused on your chosen task. Use Tools to choose what the agent normally does. An instruction saying “ask me first” is not the same as turning an unwanted tool off.

The distinction matters for publishing websites: the agent is told to ask first, but the publishing tool itself does not stop for approval (static publishing). Turn it off if you do not want that capability. Do not assume every action has the same approval gate.

Read agent approvals and limits before granting more access. Start with sample data and keep unrelated credentials out of the agent. Add access only for a task you have chosen to test.

Add repetition after a useful result.

Schedules can be created through chat or under Scheduled tasks. You can run, pause, resume, or delete them. A sleeping agent wakes when a task is due, but risky commands in scheduled tasks are blocked because nobody is there to approve them.

Before resuming a schedule, decide where fresh input will come from and who checks the output. Keep a copy of source material you cannot lose. Memory and files survive sleep and restarts, but storage is not a backup.

Count review effort as part of the cost. Kilat AI has usage limits; a more elaborate task is not automatically a better business choice. Compare the manual task with preparing, running, and checking the agent's draft.

Start with one agent and one job.

For related preparation, read about business memory and scheduled tasks. Keep your first test small enough to inspect from beginning to end. Get started.

Read next: AI agent vs chatbot: the difference, and which you need · How to create an AI agent in Telegram for your business, no coding

Built to work for you

Put your agent to work.

Start from a template and shape it to your business.

Get started