Agents

An agent is the assistant you talk to. It carries a persona (its instructions) and a model (the engine behind its answers). You choose which documents it can read when you start a chat — so the same agent can answer from different sources in different conversations.

Agents hold the persona; chats hold the sources

This is the one thing worth getting straight. An agent is not tied to a fixed set of datasets. You pick the datasets to search per conversation, in the chat itself. The agent decides how to answer (tone, model, rules); the chat decides what it reads. See Chatting for choosing sources.

The built-in assistant

Every workspace starts with one agent — RAGBot Assistant — already set as the default. It's tuned to answer from your indexed sources and cite every claim, so you can start chatting immediately without configuring anything.

This default is a system agent: you can open and read its settings, but they're locked. To create your own variation, make a new agent.

ragbot — agents / Support agent

Support agentGPT-5.4 Mini

You are a friendly support assistant. Answer only from the connected handbook and policy documents…

Created Jun 1, 2026Updated 2 days ago

Everything an agent needs: a model, a persona, and the datasets it can read.

Create an agent

Open the Agents area and choose New agent. Name it after the job it does — Support agent, Policy helper, Research assistant — then set its persona and model. Everything is editable later.

ragbot — new agent

New agent

Create agent

Name

Description

System prompt

You are a friendly support assistant. Answer only from the connected handbook and policy documents. If something isn't covered, say so and suggest who to contact.

Model

GPT-5.4 MiniBalanced

What will you mostly use it for?

GPT-5.4 MiniOur pick

Great answers at a lower cost

Temperature

012

An agent's settings: name, system prompt, model, and temperature — editable any time.

Configuration options

System prompt (instructions & persona)

The system prompt tells the agent how to behave — its tone, who it's helping, and any rules to follow. A couple of clear sentences go a long way. For example:

• "You are a friendly support assistant. Answer only from the sources in the conversation. If something isn't covered, say so and suggest contacting HR."
• "Be concise. Prefer bullet points. Always mention which document an answer came from."

The built-in assistant's prompt is a good template: answer only from the linked sources, cite every claim, and say plainly when the sources don't cover a question rather than guessing.

Model

The model is the engine behind the answers. The default (GPT-5.4 Mini) is quick and capable for most questions. Choose a stronger model when answers need more careful reasoning, and a faster one when you want speed over depth. You can switch any time.

Temperature

Temperature (0–2) controls how creative the wording is. Lower values stay close to the source text and are best for factual, document-grounded answers; higher values are looser. For most RAG use, keep it low.

Default agent

One agent in the workspace is the default — the one preselected when you start a new chat. Mark whichever agent you reach for most as the default.

How to test your agent

The best test is a real one. Start a chat with the agent, choose the sources that hold the answers, and ask the questions your readers actually ask. Check three things:

1. Is the answer correct? Click the cited sources to confirm they say what the agent claims.
2. Did it use the right documents? If it's citing the wrong material, revisit which datasets you selected for the conversation.
3. Does it admit gaps? Ask something your documents don't cover — a good agent should say it couldn't find an answer rather than invent one.

Adjust the system prompt or model, then ask again. A few rounds of this is usually all it takes to get an agent you trust.