Starts this lesson and continues through 13 more to the end of the course.

Quick chat and chat agents

Why this matters

Chat is the front door of the whole suite now — the docs describe it as the way to "ask questions, interact with your Quick resources, and accomplish tasks using multiple agentic worflows using natural language" (typo verbatim, Using Quick chat, verified 2026-08-26). When your boss asks "can our people just ask the data a question," this lesson is the answer, including the parts marketing leaves out: what it's scoped to, what it costs to trust, and where the knobs are.

The chat surface

From the same page (verified 2026-08-26):

Opens from the chat bubble icon, top-right, anywhere in Quick — collapse mode (side panel "from within all Amazon Quick contexts") or expand mode.
Context-aware: "It registers the Amazon Quick resource you're interacting with and scopes your chat automatically to the resource." Open a dashboard, and chat is about that dashboard — including auto-focus when you open a space mid-conversation.
Three data-scoping modes — this is the control that decides what an answer is built from:
- All data and apps — everything you can access: "spaces, dashboards, datasets, topics, knowledge bases, action integrations, and LLM knowledge"
- General knowledge — LLM only, no company data
- Specific data and apps — pick the space(s)/dashboard(s)/dataset(s)/topic(s) yourself
Web search is built in, per-query toggleable via the globe icon, and only exists "if it has been enabled by admin."
Datasets in chat: SPICE datasets (the familiar "2 TB and 2 billion rows" SPICE ceiling applies — Q2 lesson 2) and Direct Query against Redshift, Athena, Aurora PostgreSQL, and S3 Tables.
File uploads: up to 20 files per conversation (10 MB images/spreadsheets, 50 MB others, 665,000 parsed characters total) — remember: a space holds 10,000 files. Twenty is for ad-hoc.
Trust surface: source citations; Explanations including "Generated SQL (datasets only) – Displays the specific SQL query that produced each numerical claim"; per-user memory ("Memories are specific to you"); conversations retained "for up to 90 days" (the quota table agrees: 90 days, not adjustable).

That Generated-SQL explanation panel is the single most useful thing to teach analysts: the model's answer is auditable down to the query.

System agent vs custom agents

Working with agents (verified 2026-08-26) defines exactly two kinds:

	System chat agent ("My assistant")	Custom chat agents
Availability	"automatically available to all users by default"	shared explicitly, private until shared
Character	"an unopinionated chat agent by design", "a base planner with no inherent data or actions of its own"	persona, knowledge, and actions you choose
Data reach	spaces/topics/dashboards/KBs/actions based on user permissions	can be "Preconfigured with resources" (answers only from its configured sources) or left open

The system agent has "all chat capabilities, including file upload functionality, LLM knowledge access, toxicity and other guardrails, web search" (default-assistant, verified 2026-08-26).

Building a custom agent

Console path: Chat agents → Create chat agent (custom-agents, verified 2026-08-26). Two authoring flows: describe it in natural language and Generate, or the builder view. Builder sections, exact console terms:

AGENT PERSONA — Agent identity + Persona instructions
Communication style — presets Executive / Technical / Creative, each with Tone, Response format, Length
Reference documents — "Upto 100,000 characters of text will be extracted" (typo verbatim); formats .pdf .txt .html .md .csv .doc .docx
Knowledge sources — link spaces
Actions — link action connectors
Customization — Welcome message, Suggested prompts
Publish with Launch chat agent; private until shared. Sharing roles: Owner (edit, share, delete) / Viewer (view and use).

Reference documents vs spaces — the design decision

Agent knowledge best practices (verified 2026-08-26) draws the line:

Reference documents are in-context: "Documents that provide exact instructions and remain active in the chat agent's memory." Hard limits (Service Quotas, verified 2026-08-26, all non-adjustable): 10 documents, 50 MB total per agent, 100K characters after processing.
Spaces are retrieval: "chat agents look for answers within the space" — search first, then read what matched.

Rule of thumb the docs support: policy the agent must always follow → reference document. Library the agent should search → space. The quota table also caps resources per chat agent at 20 (not adjustable).

Embedding and the API

Agents embed in external sites: admin allowlists the domain under Manage account → Security → Manage domains, then the owner uses Share → Share via embed (custom-agents, verified 2026-08-26). AWS announced embedded chat "either through 1-click embedding or through API-based iframes for registered users" (What's New, 2025-11-24, verified 2026-08-26).

Agents are API objects — in the QuickSight API, of course:

aws quicksight create-agent | update-agent | describe-agent | list-agents |
    search-agents | delete-agent |
    describe-agent-permissions | update-agent-permissions

CreateAgent (verified 2026-08-26) is POST /accounts/{AwsAccountId}/agents — described as "Creates an agent in Amazon QuickSight", one word, in 2026 — and its shape confirms the console limits: ActionConnectors max 10, Spaces max 10, StarterPrompts max 3, plus CustomPromptInput, WelcomeMessage, and AgentLifecycle (PREVIEW | PUBLISHED).

⚠️ There is no conversation API. The API reference has agent CRUD, embedding URL generation (GenerateEmbedUrlForRegisteredUser[WithIdentity], GenerateEmbedUrlForAnonymousUser) and Q&A config ops — but nothing that sends a chat message. If someone wants programmatic chat, the supported route today is embedding, not a REST call. Chat content logging is also not in CloudTrail — it's CloudWatch vended logs (lesson 6).

Check yourself

An analyst asks "why does chat cite our wiki for a revenue number?" What setting do you look at?
When do you put a document in an agent's reference documents rather than in a linked space?
A custom agent needs 12 action connectors. What stops you, and where is that documented?
Compliance asks whether users can chat with data they aren't permitted to see through the system agent. Answer?
A developer asks for "the SendChatMessage API." What do you tell them?
How does an analyst verify a numeric claim chat just made about a dataset?

Answers

The data-scoping mode. "All data and apps" includes knowledge bases and LLM knowledge; switch to "Specific data and apps" and pin the dataset or topic if answers must come from governed data.
When it's short, mandatory context — instructions the agent must always apply. Reference documents stay in the agent's context permanently (10 files / 50 MB / 100K chars); spaces are searched at question time.
The API caps ActionConnectors at 10 per agent (CreateAgent request shape), and the quota table caps resources per chat agent at 20, not adjustable.
No — the system agent reaches spaces, topics, dashboards, KBs, and actions "based on user permissions." The caveat worth adding: files uploaded into a shared space are readable by everyone with the space (lesson 2's exception).
It doesn't exist. Agent management is API-scriptable; conversations are not. Offer embedded chat (1-click or API-based iframe for registered users) instead.
Open Explanations on the response — for dataset answers it shows the generated SQL that produced each numerical claim, plus sources and assumptions.

Teaching this section

Open by building an agent live with the natural-language flow — it takes two minutes and demystifies the whole feature.
Demo: the same question under all three scoping modes; watch the answer and citations change. Then open Explanations and read the generated SQL aloud.
Good discussion: "Which three custom agents would you ship to your org first, and what goes in reference documents vs the space for each?"
Watch for: learners treating the agent as a security boundary. Permissions ride the user, not the agent — except space files.
Timing: 25 min presented, 15 for the build-an-agent exercise.

← PreviousSpaces, knowledge bases, and the Quick Index Next →Topics, dashboard Q&A, and natural-language BI