Get started with AgentStack

AgentStack gives you one place to build an AI agent, connect it to reliable knowledge, test its answers, and make it available to customers. You do not need to configure every feature before the agent becomes useful. A strong first version has a narrow job, one trustworthy source, a clear system prompt, and a small set of questions that you can grade yourself.

The recommended path

Use this order for a new agent:

1. Create the agent. Give it a name that makes its job obvious.
2. Add one source. Start with a public website page, a file, or a Q&A pair whose contents you know well.
3. Test in Playground. Ask questions with known answers, then try an unsupported question to see how the agent handles uncertainty.
4. Improve the instructions or knowledge. Change one thing at a time so you know what helped.
5. Deploy one channel. The website widget is usually the quickest first customer surface.
6. Review real conversations. Use Activity and Analytics to find missing knowledge, confusing prompts, and failed answers.

The first-agent tutorial walks through this loop with the exact dashboard controls.

What makes an agent reliable

An agent's behavior comes from several layers:

• Sources provide the facts it can retrieve.
• The system prompt customizes its role, boundaries, tone, and uncertainty behavior alongside AgentStack's retrieval and handoff safeguards.
• The model affects reasoning quality, speed, and credit cost.
• Channels decide where customers can reach it.
• Actions let it do work, such as calling a connected tool, after you deliberately configure them.

Treat these layers separately. If an answer contains the wrong policy, fix or replace the source. If the answer is factually correct but too verbose, improve the system prompt. If the same source and prompt perform inconsistently on difficult questions, compare models.

Before you publish

Prepare a small acceptance set before exposing the agent to customers. Include at least:

• three common questions with exact expected facts;
• one ambiguous question that should trigger a clarifying question;
• one question outside the available knowledge;
• one request the agent must not fulfill;
• one question phrased differently from the source text.

The response-quality guide shows how to run this evaluation in Playground. For practical design rules, see Best practices. If you are deciding between providers or credit tiers, see Choose a model.

Go deeper when you need it

After the first version works, use the detailed guides to train the agent, install the website widget, connect other channels, add integrations and actions, and set up human handoff.

You can return to the FAQ when a first setup does not behave as expected.