You've learned what makes a system agentic, the workflow patterns, the Quick stack, and built an agent hands-on. Now design a real agentic pilot for your own team's payments workflow — and score it on autonomy, tools, guardrails, and impact.
⏱ 60 min👥 Teams of 4–5💳 Your Real Workflow🤖 Agent Design📦 Take-Home Build Kit
Capstone Overview
Across Day 2 you've gone from "why a chatbot isn't enough" to building a working Dispute Triage agent and flow in Amazon Quick. This capstone is where you turn that into a pilot proposal for your own team's workflow — designed as an agent, not just a prompt.
You'll work in cross-functional teams to fill an Agentic Pilot Canvas — a one-page framework that captures the workflow, the agent design (autonomy level + the perceive-reason-act loop + workflow pattern), the tools and actions it uses, the guardrails and human oversight, and the business impact. Then you submit it for scoring, then take home a Build Kit to stand it up in Amazon Quick.
🧭 In plain terms — what's going on here:
Now it's your turn. Your team takes a real headache from your own work — the dispute backlog you triage by hand, the reports you consolidate every Monday, the KYC checks that eat your week — and sketches on one page how an AI assistant could help: what it should do, what it must never do, and how you'd trial it safely. A scoring rubric then gives you feedback and a score. Why it matters: you walk out with a concrete, right-sized idea you could actually pitch to your manager next week — not just theory.
🎓 Lab Outcomes — what you'll be able to do afterwards
• Tell an agentic workflow apart from a one-off prompt
• Design an agent — set its autonomy level, perceive-reason-act loop, and workflow pattern
• Define guardrails and a clear human-in-the-loop boundary (assist, never decide, on fraud/dispute/credit)
• Estimate business impact and judge what's realistically buildable in Amazon Quick
• Walk away with a scored, pitch-ready pilot and a downloadable Build Kit for your own team
🔗 Bridge from Lab 1
Remember the Pilot Advisor answer you generated at the end of Lab 1? That grounded sketch of your own team's problem is the perfect seed for this canvas. Pull it up — you'll reuse it here.
Step
What you do
Duration
1
Form teams & pick an agentic workflow
5 min
2
Fill the Agentic Pilot Canvas (your own thinking — no AI)
15 min
3
✨ Reveal AI refinement → sharpen the design
15 min
4
Submit for scoring + leaderboard
10 min
5
📦 Realize it in Quick — stand up your agent from the Build Kit
15 min*
*Step 5 is "build as far as you can in the room, then take the kit home to wire it to your real data." The Build Kit is your tangible takeaway.
📦 What you take home
By the end you'll have a downloadable Build Kit for your own team's agent — paste-ready blocks for every Quick surface your pilot needs: a Space setup, a Chat Agent brief, a Flow brief, a Quick Research objective, and a Quick App prompt — in the same format you used to build the Dispute Triage agent in Lab 1. Stand up the shell in Quick during the session, then point it at your real documents back at the office.
🏆 Two-round scoring
Round 1: Submit your canvas from your own team discussion (no AI help). See how your human agent-design thinking scores. Round 2: In Step 3, reveal the refinement prompt and use AI to sharpen your agent design, then resubmit for an improved score.
The leaderboard keeps your best score. Teams that reason well about autonomy and guardrails in Round 1 start ahead.
Step 1: Pick Your Agentic Workflow
Click a workflow that resonates with your team — or pick "Custom". Each one maps to an agent pattern you saw in Modules 2–6. Your selection pre-fills the canvas context in Step 2.
💡 What makes a good agentic pilot (vs just a GenAI prompt)?
Pick a workflow that is (1) multi-step — not a single question, (2) recurring — happens weekly or daily, (3) needs the agent to take an action or use a tool (look something up, post a result, cross-reference data), and (4) has a clear human-in-the-loop safety boundary. In payments the best pilots remove the invocation, not the decision — the human still owns every fraud, dispute, and credit call.
✅ Selected:
🙋 Don't see your workflow? Pick "Custom Agentic Workflow" at the bottom and describe your own. The strongest pilots come from a process you personally run every week.
Step 2: Fill the Agentic Pilot Canvas (Your Own Thinking)
Fill this out as a team, using your own knowledge — no AI yet. This is where you reason through the agent design yourself. Short answers in each field are fine.
⚠️ Round 1: Human thinking only
Don't paste this into Quick or any AI tool yet. The goal is to capture your team's real understanding of the workflow and how you'd design an agent for it. AI-written answers score well but don't prove you understand autonomy and guardrails. AI-assisted refinement (Step 3) is Round 2 — do it after you submit.
Agentic Pilot Canvas — Fill as a team
# Agentic Pilot Canvas
## 1. Workflow & Trigger
- Workflow: [What multi-step process are we automating?]
- Who does it today: [Which roles/team?]
- Frequency: [Daily, weekly, per-event?]
- Trigger: [What kicks the agent off — a file arrives, a schedule, a question?]
- Time spent today: [Hours per occurrence]
## 2. Agent Design
- Autonomy level: [Assisted / Workflow-automation / Partial-autonomy / Full-autonomy — and why]
- Perceive: [What inputs does the agent read?]
- Reason: [What does it decide / classify / cross-reference?]
- Act: [What does it produce or do?]
- Pattern: [Prompt chaining / Routing / Parallelisation / Orchestration]
## 3. Tools & Actions
- Reads from: [Spaces, knowledge bases, documents, dashboards]
- Actions it takes: [MCP/connector calls, create/update/notify]
- Data sources: [Named systems where the data lives]
## 4. Guardrails & Human Oversight
- Agent must NEVER: [Hard limits — what it must not decide/do]
- Human reviews when: [Specific triggers — severity, value, ambiguity]
- Audit trail: [What gets logged for traceability?]
## 5. Business Impact
- Time saved: [Hours × people × frequency]
- Risk/quality improvement: [What gets safer or more consistent?]
- Success metric: [One measurable: "reduce X from Y to Z in 30 days"]
## 6. Build Plan in Quick (30 Days)
- Quick layers: [Space / Chat Agent / Flow / MCP — which ones?]
- Pilot scope: [One team, one segment, one document type]
- Go/No-Go: [What result = expand vs stop?]
✅ Tips for a strong agentic canvas
🎚️ Justify the autonomy level — match it to the risk. Fraud/dispute/credit → lower autonomy + human gate.
🔌 Name the actions — "calls list_disputes then posts to Teams" beats "the agent does stuff".
🛡️ Make guardrails specific — "never auto-decide a dispute; a Resolutions Officer reviews all escalations" not "be safe".
🧩 Name the pattern — chaining, routing, orchestration (from Module 3).
📐 Quantify impact — hours × people × frequency, plus one clear success metric.
Step 3: Refine with AI → Sharpen the Design
After you submit your Round 1 canvas and see your score, reveal the AI-assisted refinement prompt below. It's pre-filled for the workflow you picked in Step 1 — paste it into Amazon Quick Chat and the AI reviews and sharpens your agent design, asking one clarifying question at a time, then hands back a tighter canvas ready to resubmit for a higher score. You'll turn that finished canvas into a paste-ready Amazon Quick Build Kit in Step 5.
💡 Do Round 1 first
Fill the canvas from your own team discussion (Steps 1–2) and submit it in Step 4 to get your Round 1 score. Then reveal the refinement prompt below to sharpen your design with AI for Round 2.
The prompt is tailored to the workflow you selected in Step 1 — reselect any workflow to update it.
✅ Round 2: AI-Assisted Agent Design
Below is a guided refinement prompt pre-filled for your selected workflow () — the workflow, steps, pattern and starting autonomy are already filled in; the rest are [bracketed] for you to complete. Paste it into Amazon Quick Chat and it will review, improve, and ask you one clarifying question at a time as you fill each section — then hand back the finished canvas and a VP pitch.
Agent-Design Refinement Prompt (AI-Ready)
Refinement prompt — copy into Quick Chat / Research
Select a workflow in Step 1 to see the detailed, pre-filled prompt.
✅ Round 2 Checkpoint: you should now have:
• A sharpened canvas — explicit autonomy level, named tools/actions, specific guardrails
• The canvas as a copy-paste markdown block
• A 3-sentence VP elevator pitch and an agent architecture sketch
• Ready to resubmit in Step 4 for a higher score
Step 4: Submit & Score Your Agent Design
Submit your Agentic Pilot Canvas for scoring. It's evaluated on 5 agent-specific dimensions, 1–5 each (25 max). The full rubric is shown below the submit box, so nothing about the score is a black box.
📤 Submit Your Agentic Pilot Canvas
Submit as a team or individually. Resubmitting with the same name keeps your best entry.
🔍 Full scoring transparency — exactly how the score is computed
Your canvas is scored by a live LLM-as-Judge (Claude on Amazon Bedrock) against the transparent 5-dimension rubric below, and your entry joins the shared workshop leaderboard that your facilitator can review. It rewards specificity: a justified autonomy level, named tools/actions, specific guardrails, a realistic Quick build, and quantified impact. If the scoring service can't be reached, it falls back to an on-device rubric so you still get a score — that offline entry stays on your device, so resubmit from the workshop website to add it to the leaderboard.
The 5 dimensions (each scored 1–5)
Dimension
Max
What a 5 looks like
Autonomy Design
5
The autonomy level (assisted / workflow / partial / full) fits the task's risk and is justified; the perceive-reason-act loop is clear; the act-vs-escalate boundary is explicit.
Tools & Actions
5
What the agent READS (Spaces, KBs, docs) and DOES (MCP/connector actions) is named; the right workflow pattern (chaining / routing / parallelisation / orchestration) is chosen; it can finish the task end-to-end.
Guardrails & Human Oversight
5
"Must NOT" rules are specific to regulated payments work; human-in-the-loop checkpoints have triggers; there's an audit trail and a clear "never auto-decide" boundary — an auditor or compliance reviewer would be satisfied.
Feasibility
5
Honestly buildable on Amazon Quick (Spaces / Chat Agents / Flows / MCP), 30-day scope is realistic, and the data actually exists — a team could start next week.
Business Impact
5
Value is quantified (hours × people × frequency, or risk/error reduction) with one clear, measurable success metric and a Go/No-Go criterion a sponsor could validate.
How each dimension is scored (1–5)
Score
Meaning
1
Missing or completely generic — could apply to any company or any task
2
Present but vague ("the agent helps", "saves time", "it's autonomous")
3
Reasonable, but lacks specificity in autonomy, tools, or oversight
4
Strong — specific agent design and named actions, with minor gaps
Solid agent design — address the feedback and it's ready
11–15
🥉 Good Start
Right direction; sharpen autonomy, tools, or guardrails
Below 11
🔄 Needs Refinement
Add specifics — the written feedback tells you where
📈 How to score well: the rubric rewards specificity over polish. "Never auto-decide a dispute; a Resolutions Officer reviews all escalations; agent calls get_dispute then flags the 3-in-30 pattern" beats "the agent is safe and helpful" every time. Name your autonomy level, name your tools/actions, quantify the impact, and ground the guardrails in your real workflow.
Note: a rubric score is a fast, consistent coach — not an infallible authority. Two near-identical canvases can land a point apart, and the written feedback matters more than the exact number. Use it to sharpen your thinking, then trust your own judgement in the pitch.
🏆 Leaderboard
Submissions ranked by score across the whole workshop. Resubmitting under the same team name keeps your best entry.
Loading leaderboard...
Step 5: 📦 Realize It in Quick — Your Take-Home Build Kit
Your canvas is a design. This step turns it into paste-ready prompts — one per Quick surface — generated from your canvas answers. Stand up the agent shell in the room, take the kit home to point it at your real data, and use the leadership-pitch app to present the idea to your sponsor.
🎁 This is your takeaway. Paste-ready blocks for every Quick build surface — a Space setup, a Chat Agent brief, a Flow brief, a Quick Research objective, and a Quick App prompt — everything you need to recreate your pilot in Amazon Quick after the workshop.
How the kit works — driven by your canvas. Paste your completed canvas and each block is tailored to your answers (leave it blank and it uses the selected workflow's starter values). You get one paste-ready prompt per Quick build surface, each downloadable as its own .md: A Space · B Chat Agent · C Flow · D Quick Research · E Quick App (ops tool) · F Quick App (leadership pitch — a one-page app to present the idea to your sponsor) · G a step-by-step deploy guide. A and B are the core; the rest are optional depending on your pilot. Follow the checklist below.
🛠️ Generate your Build Kit
Paste your completed canvas (from Step 3/4) so the kit is tailored to your answers — or leave it blank to use the selected workflow's starter values. You'll get one paste-ready prompt per Quick surface, each downloadable as its own .md.
🛠️ Generate builds an instant, offline starter kit from your canvas (each block downloadable as its own .md). For a sharper, reasoned kit, use the “Let Amazon Quick build your kit” meta-prompt just below — Amazon Quick does the thinking, not our servers.
🤖 Or: let Amazon Quick build your kit
No server, no wait on us — Amazon Quick itself generates your kit. Copy the meta-prompt below, paste it into Amazon Quick Chat (attach your Space if you have one), then paste your completed canvas where marked. Quick returns all seven paste-ready build blocks. Same structured approach as the Step 3 design-refinement prompt, but focused purely on producing the build kit.
Meta-prompt — paste into Amazon Quick Chat, then add your canvas
You are an Amazon Quick build-kit generator for a payments & merchant-services company (AnyCompany Payments). I will give you my Agentic Pilot Canvas. Turn it into a paste-ready Amazon Quick BUILD KIT — one block per Quick build surface — that a non-developer can paste straight into Amazon Quick.
Produce EXACTLY these seven blocks, each self-contained and paste-ready (roughly 120–220 words each):
A. Space Setup — the knowledge Space: a name, a system-prompt description, and what (synthetic) documents to upload.
B. Chat Agent — persona, goal, knowledge source (the Space), named actions/tools, guardrails, and 4–5 suggested prompts.
C. Flow — numbered steps: trigger/input → pull Space data → run the chat agent → a condition / human-review gate → action.
D. Quick Research — a cited, structured report objective with deliverables (can be scheduled).
E. Quick App (Ops Tool) — an internal screen: input, a results panel with citations, a "needs human review" list, an impact strip.
F. Quick App (Leadership Pitch) — a one-page, boardroom-ready app to present the idea to leadership: exec summary, an autonomy badge, a Trigger → Perceive → Reason → Act → Human-checkpoint flow strip, guardrails & oversight, an impact big-number panel, a 30-day plan, and a "why fund this now" panel.
G. Deploy in Quick — short, numbered steps to build A→F in order (build the shell with synthetic data in the room; swap in real data later).
Rules: the company is always "AnyCompany Payments"; refer to people by ROLE only (no invented names). Assist, never decide — a qualified human owns every fraud / dispute / credit outcome, so bake a human-in-the-loop gate into B, C, and F. Synthetic data only; mask card numbers (e.g. ••••4921). Never invent a metric — write [NEED: data] where a number is missing. Label each block clearly A–G so I can copy them one at a time.
MY AGENTIC PILOT CANVAS:
[PASTE YOUR COMPLETED CANVAS HERE]
✅ Why this is the win-win: the reasoning happens inside Amazon Quick (the tool you're here to learn) — nothing is sent to any workshop server, and there's no endpoint or infrastructure to secure.
Stand It Up in Quick (same moves as Lab 1)
If you have time in the room, build the agent shell now. Otherwise this is your at-the-office checklist.
① Create the Space — Quick → Spaces → Create space. Use the Space Setup from your kit as the name + description. Upload a few representative documents (sample now, real set later).
② Build the Chat Agent — Quick → Chat agents → Create chat agent → paste your kit's Chat Agent Brief → Generate → confirm the Space is attached → Update preview → test → Launch chat agent.
③ (Optional) Add the Flow — Quick → Flows → Create flow → Natural language → paste your kit's Flow Brief → Generate → run it once → schedule or share with your team as Viewer.
④ (Optional) Add Quick Research — Quick → Research → paste your kit's Research Objective → run it → if it should recur, save it as a scheduled research agent. Best for open-ended, cited investigation workflows (vendor risk, market/scheme landscape).
⑤ (Optional) Quick App — operations tool — Quick → Apps → describe your app → paste block E → Generate → iterate in the live preview → Publish. A clickable front-end screen for running the workflow.
⑥ Quick App — leadership pitch — Quick → Apps → paste block F → Generate → refine → Publish and share the link with your sponsor. A one-page, boardroom-ready app to present the agent idea to leadership.
⚠️ Data reality check. In the room you likely won't have your real team documents loaded — that's fine. Build the agent shell now to prove the design works, then back at the office swap in your actual files (and add an MCP integration if you need live system data). The kit is the recipe; your data is the ingredients. And remember: all sample data here is synthetic (masked PANs like ••••4921, synthetic IDs) — never load real cardholder data into a pilot.
🎓 What You Just Did — and Where Next
In one capstone you took everything from Day 2 — the LLM-to-agent shift, the perceive-reason-act loop, the four workflow patterns, the Quick automation stack, and the hands-on build — and turned it into a fundable pilot for your own team's payments workflow.
📌 Take your canvas home. It's not a workshop artifact — it's a deployment blueprint. The build plan maps directly onto Quick: a Space for knowledge, a Chat Agent for the reasoning, a Flow to automate it, an MCP integration when you need live system data. You can prototype it in an afternoon.