> ## Documentation Index
> Fetch the complete documentation index at: https://docs.linkiasoft.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Build an AI agent, step by step
> The six steps — Identity, Guardrails, Knowledge, Training, Review & Compile, Assign — and what actually matters in each.
**AI Agents → Create Agent.** Six steps. You can move back and forth; nothing goes live
until you assign channels in the last one.
## 1. Identity
Who the agent is and how it talks.
| Field | What it does |
| -------------------------------- | -------------------------------------------------------------------------------------------- |
| **Agent name** | Internal label. Required. |
| **Language** | The language it replies in. |
| **Tone** | Professional, Friendly, Casual, Formal or Empathetic. |
| **Brief description** | A short summary of your business. Used to ground the training questions generated in step 4. |
| **Instructions (system prompt)** | The real content — who you are, what you sell, how to greet, what to never do. |
| **Greeting message** | Sent when a new conversation starts. |
| **Fallback message** | Sent when it cannot answer. |
| **Tools** | Which [tools](/help/ai/tools) this agent may call. |
**Instructions** is the field that decides whether the agent is any good. Write it as if
briefing a new employee on their first day: who we are, what we sell, our delivery times,
what to say when we are out of stock, when to fetch a human. Vague instructions produce
vague replies.
## 2. Guardrails
Limits on what it may say. Covered in detail in [Guardrails](/help/ai/guardrails).
## 3. Knowledge
Two ways to give it facts:
PDF, TXT, DOCX, CSV or MD, up to 10MB each. Price lists, policies, FAQs.
Enter your site URL. We crawl it, extract the content and summarize it into the agent's
grounding. This can take up to a minute.
## 4. Training
This step is easy to skim past and shouldn't be.
Press **Find Gap Questions**. The system generates the hard, business-specific questions
the agent **cannot confidently answer** — the ones needing knowledge only you have. For
each, it shows its own attempt and its confidence.
Your job is to supply the correct answer. Pick a suggested answer, edit one, or write your
own. You can discard questions that are not relevant.
This is deliberately not "here are questions it answers well". Those need no work. The
value is entirely in the gaps — the questions where the agent would otherwise invent
something plausible.
Generation takes a minute or two and continues in the background if you leave the step
open.
## 5. Review & Compile
Compiling assembles everything above into the agent's working configuration and produces
an **eval score** by re-asking its own training questions.
Re-compile any time you change instructions, knowledge or training answers. Until you do,
your edits are not in effect.
## 6. Assign
Choose which channels it replies on. **Nothing happens until this step** — an agent with
no channels is inert.
A channel can only be assigned to one agent at a time; a channel already taken shows as
**Assigned to another agent**.
## After it is live
* **Overview → Edit** to change any of the above, then re-compile.
* **Set as Default** from the agent list.
* Set **Inactive** to stop it without deleting it.
* Watch its first real conversations in the inbox. The eval score is not a substitute.
## Related
Prohibited topics, never-rules, PII handling, escalation.
Giving it the ability to act, not just talk.