Agents

An agent is a reusable assistant with its own instructions, tools, knowledge and model. Build one once, and anyone you share it with gets the same behaviour every time.

A prompt tells a model what to do this time. An agent tells it who to be every time: the job, the rules, the documents to trust and the tools it may use. Your team then chats with the agent instead of re-explaining the job in every conversation.

The Support Triage agent answering a billing ticket. Every claim links to the Help Center document it came from.

What an agent is made of

Part	What it does	Where you set it
System prompt	The job description: role, steps, tone, what never to do.	Studio tab
Model	The model that runs the agent. Pick one that suits the job.	Studio tab, AI model
Tools	Web search, remote MCP servers and inline image generation.	Studio tab, Tools
Files	Reference files attached to the agent itself.	Files tab
Knowledge	Knowledge bases the agent searches and cites.	Knowledge tab
Details and sharing	Name, description, handle and who can use it.	Details and Share tabs

Create an agent

Open Agents and choose New agent

Go to Agents in the left sidebar, then select New agent (or Create agent if it is your first).
Name it and say what it is for

Give it a name your colleagues will recognise, such as Support Triage, and a one-line description. The description appears on the agent’s card, so write it for the person deciding whether to use it.

Write the system prompt

Describe the job the way you would brief a new colleague: what to do, in what order, in what tone, and where the limits are. Numbered steps work well.

You are the first-line support agent for Northwind Analytics.

For every ticket:
1. Classify it: billing, bug, how-to, feature request or account access.
2. Rate urgency 1-3 (3 = a customer cannot work).
3. Draft a reply in our voice: warm, direct, no jargon, max 120 words.
4. If it is a bug, list the reproduction steps you still need.

Use the Help Center knowledge base as the source of truth.
Never promise dates or refunds - hand those to a human.

Choose the model and tools

Pick the AI model on the right. Then switch on the tools the job needs: Web Search for live, sourced results; Remote MCP to reach your own MCP servers; Image Generation to create images inline.
Save

Select Create agent. It now appears on the Agents page, ready to chat with.

Give it your documents

Agents answer best when they can check their facts. Open the agent’s Knowledge tab and tick the knowledge bases it should use. Replies then carry inline citations that point to the exact file each claim came from.

The Knowledge tab of an agent with the Northwind Help Center knowledge base available to attach

World-Wide Knowledge lets the agent also use general knowledge beyond your documents. Leave it off when answers must come only from your sources.
Use all knowledge bases attaches every knowledge base you can access, including ones created later.
Or tick individual knowledge bases for tight control.

Knowledge bases are covered in depth in Knowledge bases.

Chat with an agent

Open the agent and select Chat, or pick it from the agent picker in any chat composer. The composer shows the agent’s name where the model normally is, so it is always clear who is answering.

A reply from the Support Triage agent with citations to Northwind Refund Policy.md and Northwind Onboarding Guide.md — Citations appear as highlighted file names right after the sentence they support.

Every conversation with an agent is kept under History on the agent’s page, so you can review how it has been answering.

Use the Share tab to make the agent available to colleagues or your whole workspace. Everyone then chats with the same instructions, tools and knowledge, so the answers stay consistent no matter who asks.

Good first agents

Support triage

Classifies tickets, drafts replies from your help center, flags anything urgent. Attach your policies as a knowledge base.

Market researcher

Web search on. Returns a one-page, sourced brief on any company or market.

Brand voice editor

Rewrites drafts to your style guide and explains what it changed.

Sales assistant

With the CRM connected, looks up accounts and drafts follow-ups. See Integrations.

Agents can act, not just answer

Everything a chat can do, an agent can do as part of its job: query your connected systems, make a PDF or Word document, generate images, and email a colleague by name - StickyPrompts looks teammates up in your workspace, so “send the draft to Sarah for approval” needs no address. See Every tool, any model.

Automate an agent’s work

When you find yourself asking an agent the same thing on a schedule - every Monday, every morning - turn it into a workflow. A workflow runs fixed instructions on a model automatically and keeps a history of every run.

Use agents from the API or MCP

Agents are available to your own code too. List them with the REST API (GET /agents) or the MCP tools list_agents and get_agent, and pass an agent to a conversation. See Agents and workflows.