This article is auto-synced from its in-app version in Tai.
Custom agents let you build an AI assistant tailored to your company's workflows — combining skills, connectors, and knowledge to answer questions or automate tasks specific to your business.
When you'd use this
Build a custom agent when the built-in agents don't cover your use case. For example: create an agent that knows your cancellation policy (as a knowledge item), can query your booking system (via a connector), and follows your specific support procedures (via a skill) — and make it available to your team in a chat.
Steps
1. Plan your agent
Before creating the agent, identify:
What questions should it answer? — this shapes the system prompt.
What data does it need? — if it queries live systems, you need a connector.
What procedures should it follow? — write these as a skill or attach them as knowledge.
Who will use it? — pick a name and welcome message that fit your team's language.
2. Set up the building blocks
Create (or identify) the building blocks your agent will use before wiring them together.
Skills — add a skill if your agent should follow a specific procedure or have access to a reusable capability. See Using skills.
Connectors — add a connector if your agent needs to query a live system, such as your booking platform or an internal API. See Using connectors.
Knowledge — create a knowledge item if your agent should draw on specific documents, such as a cancellation policy or product FAQ. See Using knowledge.
3. Create the agent
In the Workbench sidebar, click Agents (under the AI Platform section).
Click Create Agent in the top-right corner.
Fill in the General tab:
Agent ID — a unique identifier in
yourcompany/agent-nameformat (e.g.acmetours/pre-departure-check).Name — the display name your team will see.
Description — what the agent does. Write this precisely — it also helps the system route requests to the right agent.
Skills — select the skills you prepared in step 2.
Connectors — select any connectors the agent needs.
Knowledge — attach any knowledge items.
Open the Behavior tab:
System prompt — describe the agent's role, what it should and should not do, and any specific instructions. Plain English works well.
Welcome message — the first message shown when a user opens a chat. Keep it short and action-oriented.
Open the Channels tab and enable Streaming API to make the agent available in chat.
Click Create Agent.
4. Test the agent
Open the agent's detail page.
Click Test Chat (top-right) to open a chat session with your new agent.
Try your key use cases and follow-up questions.
If the agent's answers are off, refine the system prompt on the Behavior tab and save.
5. Iterate
If the agent is missing information, attach additional knowledge items or extend a skill's content.
If it needs access to live data, check that the required connector is assigned and properly configured.
Changes to skills, connectors, or knowledge take effect on the next chat session.
Tips and limits
Start with a simple, focused system prompt — a precise description of the agent's role leads to more reliable answers than a long list of rules.
A skill's Description controls when the agent decides to use it — write it carefully.
If a skill requires a connector, that connector must also be assigned to the agent.
An agent can only access the skills, connectors, and knowledge items explicitly assigned to it — there is no implicit access to other resources.
You can copy an existing company agent (or one of your own) as a starting point: open its detail page and click Copy, then customise the copy. The copy is Personal by default; Company Admins can copy into the Company tier instead. (Built-in agents are Nezasa-owned and can't be copied.)
Related