Quickstart

Quickstart

This walkthrough takes you from an empty workspace to a live support agent answering real questions. It assumes your cove1 workspace has already been provisioned; if it hasn't, book a demo and we'll set it up with you — see Deployment for how rollouts work.

The path is four steps: connect a channel, give the agent knowledge, design its behavior on the canvas, and go live. A fifth section shows what handoff looks like once conversations start arriving.

1. Connect a channel

A channel is how customers reach the agent. Start with one and add more later — the same agent serves all of them.

1. Open the console and go to Channels.
2. Choose a channel type. The web chat widget is the fastest to test, so start there.
3. Create the channel. cove1 generates an agent ID and a widget snippet for it.
4. Add the snippet to any page you want the widget to appear on:

<script src="https://cdn.cove1.com/widget.js" data-agent="YOUR_AGENT_ID" defer></script>

Replace YOUR_AGENT_ID with the ID from your channel. Load the page and the widget should appear in the corner.

2. Import knowledge

An agent with no knowledge can only reason from the model's general training. Ground it in your own material so it answers with your facts.

1. Go to the Knowledge base and create a collection.
2. Upload your sources — product docs, help-center articles, policies, internal references.
3. Wait for the collection to finish indexing. cove1 splits and embeds the content so the agent can retrieve the relevant passages at answer time.
4. Test a query against the collection to confirm the right passages come back before you wire it into the agent.

Keep collections focused. A collection per product or per topic is easier to maintain than one collection holding everything, and it lets you scope which knowledge a given agent can reach.

3. Design the agent on the canvas

The canvas defines what the agent actually does with an incoming message.

1. Create a new agent and open its canvas.
2. Set the agent's role and tone — who it is, what it helps with, and the boundaries of what it should attempt.
3. Add a knowledge step and point it at the collection from step 2, so the agent retrieves before it answers.
4. Add a handoff step and define when it fires: an explicit request for a human, low confidence, or a topic you would rather a person handle.
5. Attach the agent to the channel you created in step 1.

Save the canvas. Everything the agent does now traces back to a step you can see and edit, rather than a hidden prompt.

4. Go live

1. Use the built-in preview to run test conversations against the finished canvas.
2. Try the questions you expect most often, plus a few the agent should refuse or escalate.
3. When the answers look right, publish the agent.
4. Confirm the live channel is now served by the published version.

The widget from step 1 is already pointing at this agent, so once it is published, real conversations start flowing in.

5. Watch the handoff

Handoff is what keeps an AI agent safe to put in front of customers. When a conversation hits the condition you set on the canvas, the agent stops and routes the thread to a human.

1. Open the console inbox and keep it in view while you test.
2. In the widget, ask something that should trigger handoff — for example, request a person directly.
3. The conversation appears in the human queue with the full history attached, so whoever picks it up has the entire context.
4. A human replies, resolves the issue, and can hand the conversation back to the agent when it is done.

From here, refine the canvas as you learn from real conversations: tighten the handoff conditions, add knowledge for the questions that come up most, and connect the next channel.