A tour of the place you will build: four building blocks, one knowledge boundary, and the difference between facts and rules.
01 · The shape
Four building blocks on one engine
Amazon Quick is a no-code place to build AI assistants. It runs on Amazon Bedrock, which you never open yourself. Click any box, or watch the tour.
What you supply What keeps it honest Where you work Separate tier: ask DNA / IT The engine, not yours to open
Real work touches two or three of these boxes, never all twelve. Most people stop at Chat Agents.
The same set as a stack, and why this tour comes first
Why this tour comes first. In F3 you build a working assistant in about an hour. That goes better if you already know what a Space is, why an agent has five parts, and where a rule must live to survive word for word.
The diagram shows how the pieces flow. This is the same set stacked, with context at the bottom and capability at the top. These are four building blocks, not a list of menu items. Click a layer.
Read it top-down. The top gives capability (reach other systems, run unattended). The bottom gives context (the documents that make an answer true). Either one alone is useless.
Build it bottom-up. Start with a Space. Add a Chat Agent that reads it. Once you trust the agent, put it on a schedule. Connect other systems last.
Most people stop at layer 2, and that is a complete outcome. This workshop spends its time on layers 1 and 2. You will see layers 3 and 4 so you recognise them and know what to ask for.
Looking ahead · AgentSea
Where AgentSea Is Heading Next
02 · Layer 1
A Space is the shelf your agent reads from
Put a document in a Space once and it stays there for every chat and agent. Click through creating one.
Create your space
CancelCreate
HR Guidelines & Circulars starts empty
Nothing to read yet
Staff leave guideline
Latest circular
Compliance register
Sample data in training is synthetic. What Quick may hold at Synapxe: ask DNA / IT.
Common myth: the Space description is the system prompt
MythWrite your ground rules in the description and every chat obeys them.
RealityIt is optional metadata: a short label for people, up to 512 characters. It does not steer the agent. The field asks “What are you trying to achieve?”, which invites rules. Do not put them there.
InsteadBehaviour comes from the agent’s instructions and reference documents. See .
A Space controls what the agent can read, not how it behaves.
Tell me more: what goes in, sharing, and why bother
Without a Space
Every conversation starts from nothing
You re-attach the same documents each time
Two colleagues asking the same question can get different answers, because they attached different files
With a Space
Documents are already indexed and searchable
Every agent on the Space shares the same grounding
Update one file and every agent reasons with the new version, with no agent edit
Four things you can put in. Everything goes in through + Add. A file from your own machine comes in through Upload files; external sources are attached through Add knowledge. Most teams only use the first.
Files: PDF, Word, Excel, PowerPoint, Markdown, images. Thousands per Space, indexed automatically. Folder structure is not preserved, so put the context in the filename.
Knowledge bases: a SharePoint site, an S3 bucket, a Google Drive folder, Confluence. Use when the content already lives elsewhere. What is available depends on what your administrator enabled.
Dashboards and data: Quick Sight dashboards and structured datasets. The agent can read the numbers and chart them back.
Actions: connectors to other systems, so an agent can send a message or raise a ticket. See Beyond One Agent.
A new Space opens on an empty knowledge panel: naming it gives it nothing to read, so adding files is a separate step.
Shared or private. A Space can be shared with your team or kept to yourself, and it inherits Quick’s existing permissions. Anything in a shared Space is visible to everyone it is shared with.
Situation
Same Space or new?
Another circular for the same policy area
Same Space
A different function: HR, Finance, Governance
New Space (different audience, different documents)
One active tender or procurement exercise
New Space, one per exercise
Your own draft notes
New private Space
03 · Layer 2
An agent is five parts you can fix one at a time
A Chat Agent is an assistant you define once and run again and again. Describe it in plain English, press Generate, then review its five parts.
Wrong answer? Check the knowledge source first, then the goal, and the persona last.
Tell me more: the rule of three, and building one
The rule of three. If you have typed roughly the same long instruction three times this month, that is an agent. The version you type next week will not match the one you typed last week, so your output quietly drifts.
Re-typing the instruction
Minutes spent wrestling the prompt every time
Output shape changes between runs
A colleague writes their own version, now there are two
A saved agent
The same shape comes back every time
Guardrails hold, because they are part of the definition
Improve it once and everyone inherits the improvement
From blank page to working agent. About ten to fifteen minutes the first time.
Step
What you do
What Quick does
1
Left menu, Chat agents, create a new one
Opens a way to start
2
Write a plain-English brief: who it is, what it produces, what it reads, what it must never do
—
3
Press Generate
Fills in the five parts and attaches the matching Space
4
Review the fields. Add reference documents for rules that must stay exact
—
5
Test in the preview pane: a normal case, an edge case, and one you expect it to refuse
Answers beside the configuration
6
Press Launch chat agent
Saves and publishes. Until then, nothing is saved
Test the refusal, not just the answer. If the agent should decline something, ask for that thing on purpose and check it holds the line. A guardrail you have never tested is a guardrail you do not have.
Not private. The agent’s name and description are visible to everyone you share it with. Keep vendor names, figures and personal details out.
04 · The idea F3 turns on
Facts to read, rules to follow
Quick may reword what you type into the instructions. AWS advises putting guidelines that must stay unchanged in reference documents. Watch one rule in both homes.
In the instructions field
What you typed
Quick rewrites it with AI
What the agent may end up working from
Same intent, different words. The exact refusal sentence is gone.
In a reference document
What you uploaded
Held word for word
What the agent works from
Preserved exactly. You can show an auditor where the wording came from.
Where does each belong? Click one
Try it
Pick an item to see which home it belongs in.
Facts in the Space. Behaviour in the instructions. Exact wording in a reference document.
Tell me more: the three places, limits, and why F3 is built this way
Where you put it
What it does
The catch
Persona and instructions (typed into the agent)
Sets high-level behaviour, tone and goals
Quick rewrites it with AI. Your exact wording is not preserved
Reference documents (files attached to the agent)
Held in the agent’s permanent memory for every conversation
Wording is preserved exactly. Limited to 10 files and 50MB
The Space (knowledge source)
Searchable knowledge the agent queries when it answers
Retrieval only. It grounds what the agent can read, never how it behaves
Instructions suit
Tone and style
The overall shape of the output
“Escalate when you are not sure”
A general description of the approach
Reference documents suit
Clause text that must be quoted, not summarised
A refusal or escalation script with fixed wording
Exact thresholds and entitlement tables
A checklist that must run in order
Any template an auditor will compare against
Limits to plan around: 10 reference documents per agent, 50MB total, 100,000 characters after extraction, thousands of files in a Space. Ten files is not room for a whole policy library. That is what the Space is for. Reference documents are for the short, exact things.
The test to apply. If someone could reasonably ask “show me the exact wording the agent was given”, it belongs in a reference document. If the answer is “roughly this, however you would phrase it”, the instructions field is fine.
Why F3 is built this way. In F3 you ground an assistant in one staff guideline, then get it to admit when the guideline is silent. Ask something the guideline covers and it quotes the clause. Ask something adjacent but not covered, and a poorly grounded assistant invents a confident answer. Same model, same question. The difference is where you put the rule.
05 · Layers 3 and 4
When one agent is not enough
Tick anything true of the task you have in mind. The best fit updates as you go.
Flows are available on all plans. Quick Automate is Enterprise tier and instructor demo only in this workshop; whether Synapxe has it is not confirmed, so ask DNA / IT.
These layers stack. Get the agent right first, because a Flow only automates whatever the agent does.
Tell me more: the four capabilities and how they combine
Available now
Quick Flows
A multi-step workflow that runs on a trigger, with steps described in plain language.
Triggers: a schedule, an inbound email, or a Quick event
A step can invoke a Chat Agent you already built
Run history is kept for 30 days. Save outputs to a Space for longer
Your team keeps running the published version while you edit a draft
Available now
Quick Research
You write an objective. The agent decides how to investigate and returns a cited report.
Fast: 4–7 minutes. Deep: 20–40 minutes. Set it going and come back
Shows its plan first so you can add or rule out topics
Inline citations, and a way to see the reasoning behind a claim
The executive summary tier carries no citations
Depends on your admin
Connectors and Actions
How Quick reaches the systems you already use. Two kinds.
Knowledge bases let an agent read: SharePoint, S3, Drive, Confluence
Action connectors let an agent do: Outlook, Teams, Jira, ServiceNow
Your administrator decides which exist
Connect the reading before the doing
Subject to tier
Quick Automate
The enterprise tier above Flows, for regulated processes that run unattended at scale.
Adds human approval gates, formal versioning and monitoring
Owned by IT or a centre of excellence
Enterprise tier only. Whether Synapxe has it is not confirmed: ask DNA / IT
Rule of thumb: if your team owns it end to end, it is a Flow. If the organisation owns it, it is an Automate conversation
They compose. A Flow wakes on a schedule, runs the Chat Agent you built, which reads from your Space and posts the result through a connector, before anyone opens a laptop.
Example: a weekly compliance digest. Every Monday a Flow runs your Policy and Circular agent against the HR Guidelines & Circulars Space, asks what changed against the compliance register, and drops a short summary into the governance team’s inbox. Nobody built anything new. The Flow just stopped waiting to be asked.
06 · Before you start
Six snags that cost people twenty minutes
None are hard. Click one to see the fix, and you will recognise it when it happens in F3.
Simplified map of the four building blocks. Not the literal menu.
Spaces layer 1
Chat agents layer 2
Flows layer 3
Connectors layer 4
Automations enterprise tier
The real left menu is longer, and Connectors is not one of its top-level items. Automations only appears on the enterprise tier.
Tell me more: the whole page in six lines, and what comes next
Amazon Quick
AWS’s no-code platform for building AI assistants. Amazon Bedrock is the engine underneath; you never touch it directly
A Space
The persistent knowledge boundary: what your agent is allowed to read. Its description is a label, not a system prompt
A Chat Agent
Five parts: persona, goal and instructions, knowledge source, actions, suggested prompts. Generated from a plain-English brief
Facts vs rules
Facts in the Space. Behaviour in the instructions. Anything that must survive word for word in a reference document
Flows and Research
Work that runs without you, and investigations the agent plans itself. Both available today
Quick Automate
The enterprise tier for regulated, organisation-wide processes. An IT conversation, subject to your subscription
One more, harmless but distracting. After an answer, Quick often offers suggested follow-up buttons such as “Export as DOCX”. They are shortcuts, not steps. During the labs, type the prompts from the workbook so the whole room stays in sync.
F1 · Next
From Chatbots to Agents
What makes something an agent rather than a chatbot, and what you can hand over to it.
F2 · Prompt
The CRAFT of a Prompt
Five slots (Context, Role, Action, Format, Trust) for writing the instructions an agent runs on.
F3 · Build
The Guideline Assistant
Set up a Space, generate an agent, ground it in one guideline, and get it to admit what the guideline does not cover.
F4 · Design
Agent Design Canvas
Design an agent for a workflow from your own team: the one-pager you take back.
The lab environment only runs for 90 minutes once started, so sign in during the break before F3.