Build your agent
Configure an AgentStack agent’s knowledge, retrieval, behavior, appearance, and secure integrations.
Building an AgentStack agent is an iterative loop: add authoritative knowledge, decide how the agent should retrieve it, test the result, and save only the changes that pass your evaluation. The agent sidebar groups the main building controls under Sources, Playground, and Settings.
Start with knowledge
The Sources → Content screen can ingest a website, a single page, files, Notion pages, and explicit Q&A pairs. They are not interchangeable. A website crawl is useful for a public documentation area; a Q&A pair is better for one canonical correction. See Sources overview before choosing a format.
Wait for an asynchronous source to finish before grading answers. A submitted crawl or upload is not evidence that its material is retrievable. Keep obsolete sources out of the knowledge base so an old policy does not compete with its replacement.
Set retrieval deliberately
Sources → Retrieval controls the confidence guidance and search mode. Standard retrieval is the default. Hybrid, Deep Search, and Auto are useful only when a fixed test set shows they improve the questions your customers actually ask. The retrieval guide covers the modes and their plan availability.
Test before saving
Playground lets you change the model and system prompt, use Apply & Test, and compare several configurations without saving each experiment. Save the winning model and prompt only after they pass representative supported, unsupported, and ambiguous questions. Read Use Playground for the distinction between test-only changes and saved settings.
Finish the customer experience
Under Settings, configure the agent’s general behavior, conversation opening, widget appearance, public access controls, custom domains, and webhooks. These settings have different permission requirements and save behavior, so use their focused guides:
Use one controlled change at a time. A reliable agent comes from knowing whether a result changed because of the source, prompt, model, retrieval mode, or customer surface.