📋 What We Covered Today
| Session | Topic | Key Takeaway |
|---|
| M1 | From LLMs to Agents | Fallacy of composition — what works for parts won't work for the whole. Agents need perception, reasoning, and action. |
| M2 | Exploring Agentic AI | 4 agent types: Workflow (predictable), Autonomous (self-directed), Hybrid (balanced), Multi-Agent (collaborative) |
| M3 | Workflow Patterns ⭐ | Chaining, Parallelization, Routing, Orchestration — the 4 patterns that power every agent workflow |
| M4 | Amazon Quick Suite | Quick Index + Research + Sight + Flows + Automate — enterprise AI for business intelligence |
| — | The Kiro Automation Stack | Steering (global rules) + Skills (on-demand expertise) + Hooks (automated triggers) + MCP (data connections) |
| — | MCP & Agent Design Canvas | Connect AI to live data sources. Design agents your tech team can build. |
🔄 The 4 Workflow Patterns
1. Prompt Chaining
Sequential steps where each output feeds the next. Trades speed for accuracy — each step is testable and debuggable independently.
AnyCompany example: Invoice → Extract data → Validate vs PO → Approve/Escalate → Generate report
2. Parallelization
Multiple tasks run simultaneously, then results are aggregated. Faster than sequential and provides multiple perspectives.
AnyCompany example: Merchant risk — Risk Analyst + Product Lead + Compliance Officer assess simultaneously → Unified GREEN/AMBER/RED rating
3. Routing
A classifier directs input to the right processing pipeline based on content type. Each path is optimized for its category.
AnyCompany example: Incoming document → Classify (invoice/receipt/contract) → Route to appropriate workflow
4. Orchestration
The most complex pattern — dynamically manages workflows, spawns subtasks based on conditions, and includes human-in-the-loop at configurable thresholds.
AnyCompany example: Credit approval — <$10K auto-approve, $10K-$50K AI+human, >$50K human-only. Spawns investigation if fraud flag detected.
Key insight: Most real workflows combine 2-3 patterns. A PayLater application uses routing (classify type) + parallelization (simultaneous checks) + chaining (sequential assessment). Your Day 2 prompt templates power each step.
🧪 Hands-On Labs Completed
| Lab | What You Built |
|---|
| Lab 8: SKILL.md | Converted Day 2 prompt template into a production-ready SKILL.md with steering file (global rules), hook (automated trigger), and parallelization (multi-perspective assessment) |
| Lab 9: MCP Fraud Detection | Connected Kiro to a merchant database via MCP — queried live transaction data in plain English, built a fraud detection skill that reads from the database instead of pasted CSV files |
| Lab 10: Agent Design Canvas | Designed a complete agent workflow for a real team process — chose workflow patterns, defined data sources, guardrails, escalation rules, and business impact. Scored by LLM-as-Judge. |
| Lab 11: Agentic Invoice Processing | Built a multi-step agentic pipeline — routing invoices by type, parallel compliance reviews, full end-to-end orchestration with Strands Agents SDK |
Capstone: Agent Design Canvas — team presentations with peer feedback and LLM-as-Judge leaderboard scoring.
⚡ Key Concepts Reference
Agentic Loop
Observe → Plan → Act → Reflect — continuous cycle, not linear
Perception → Reasoning → Action
The 3 components that make an AI system "agentic"
Fallacy of Composition
What works for parts won't work for the whole — why chatbots aren't enough
Autonomy Levels (L1-L4)
L1: Rule-based → L2: AI-enhanced → L3: Autonomous → L4: Multi-agent
MCP (Model Context Protocol)
Standardized way to connect AI to databases, APIs, and tools
A2A (Agent-to-Agent)
Protocol for agents to discover and communicate with each other
Human-in-the-Loop
Configurable thresholds where humans review AI decisions
Agent Design Canvas
Structured one-page document to hand to your tech team for implementation
🏗️ The Kiro Automation Stack
| Layer | What It Does | Who Owns It | File Location |
|---|
| Steering | Global rules that apply to every AI interaction — compliance standards, formatting, brand voice | Team leads / policy owners | .kiro/steering/*.md |
| Skills | On-demand expertise that activates when a specific task is mentioned — your prompt templates as reusable skills | Domain experts / Citizen Developers | .kiro/skills/*/SKILL.md |
| Hooks | Automated triggers — run actions when files change, before/after tool use, or on manual trigger | Process owners | .kiro/hooks/*.json |
| MCP | Data connections — connect AI to databases, APIs, and external tools via standardized protocol | Tech team | .kiro/settings/mcp.json |
The key distinction: Skills = your domain knowledge (instructions you write). MCP = technical connections (your tech team sets up). You own the "what" and "how to assess." They own the "where the data lives."
📦 What You Take Home (All 3 Days)
✍️
Prompt Template
Day 2 — Reusable template with {{variables}}
🛠️
SKILL.md
Day 3 — Auto-activating skill file
📜
Steering File
Day 3 — Global team rules
⚡
Hook Config
Day 3 — Automated trigger
🤖
Agent Design Canvas
Day 3 — Workflow design for your tech team
🎯 Interactive Explainers (Self-Paced Reference)
🎓 Workshop Complete
3-Day Journey: Foundations → Skills → Automation
- Day 1 — GenAI Essentials: What GenAI is, how it works (tokens, embeddings, transformers), responsible AI, and security foundations
- Day 2 — Prompt Engineering: The 4 Pillars, Chain-of-Thought, personas, structured outputs, RAG grounding, LLM-as-Judge, reusable templates
- Day 3 — Agentic AI: Agent types, workflow patterns, the Kiro automation stack, MCP connections, and designing agents for your team
🚀 What's Next: Take the Agent Design Canvas you created today and share it with your tech team. The SKILL.md and steering files you built are ready to use in Kiro and Claude Cowork today. Start with one workflow, prove the value, then expand.
💬 Participant Q&A — Day 3 Discussions
Questions raised during Day 3, answered with context and practical guidance.
Q19 Can I create my own MCP server to connect to Google Slides?
Yes — and this is exactly how MCP is designed to work. MCP is an open standard, so anyone can build a server that exposes tools for any application.
For Google Slides specifically, there are two approaches:
- Use an existing community MCP server: Search the MCP servers directory — there are community-built servers for Google Workspace (Docs, Sheets, Slides) that you can install and configure.
- Build your own: If you need custom functionality, you can build an MCP server using the MCP SDK (available in Python and TypeScript). Your server would use the Google Slides API to create, read, and modify presentations.
| Approach | Effort | Best for |
|---|
| Community server | Minutes to set up | Standard operations (create slides, add text, read content) |
| Custom server | 1-2 days for a developer | Custom workflows (branded templates, specific formatting, batch operations) |
The pattern: This is the same MCP concept from Lab 9 — instead of connecting to a merchant database, you'd connect to Google Slides. The AI agent uses the same tool-calling mechanism; only the MCP server (and what it connects to) changes. Your tech team would set this up; you'd use it by saying "create a presentation about Q3 results."
Q20 What is the difference between OAuth and API Key authentication in AI tools?
Both are ways to prove identity when connecting to services, but they work very differently:
| Aspect | API Key | OAuth |
|---|
| What it is | A static secret string (like a password) | A protocol that grants temporary, scoped access tokens |
| Security | If leaked, full access until manually revoked | Tokens expire automatically (minutes to hours) |
| Permissions | Usually all-or-nothing access | Scoped — can limit to "read only" or specific resources |
| User identity | Identifies the application, not the user | Identifies both the application AND the specific user |
| Best for | Server-to-server calls, simple integrations, development | User-facing apps, accessing personal data (Google, Microsoft) |
| Example | Bedrock API calls via AWS credentials | "Sign in with Google" to access your Drive files |
For AnyCompany: When your tech team sets up MCP connections, they'll use OAuth for services that access user-specific data (Google Workspace, Microsoft 365) and API keys for internal services (databases, internal APIs). The choice affects security posture — OAuth is preferred for anything touching personal or sensitive data because tokens expire and permissions are scoped.
Q21 How do Kiro skills translate to Claude Cowork? Can I use what I built today in both?
Yes — the core content is portable. The prompt template, domain knowledge, and business rules you wrote are plain text that works in any AI tool. The automation layer (hooks, MCP) is Kiro-specific, but the skill content transfers directly.
| What you built | In Kiro | In Claude Cowork |
|---|
| SKILL.md content | Auto-activates as a skill | Paste into a Project's custom instructions |
| Steering rules | Always-on in .kiro/steering/ | Add to Project instructions or paste at start of conversation |
| Prompt template | Part of SKILL.md | Save as a reusable prompt in your Project |
| Hooks | Automated triggers in .kiro/hooks/ | Not available — manual trigger only |
| MCP connections | Configured in .kiro/settings/mcp.json | Not available — use file uploads instead |
Translation guide: Copy the content of your SKILL.md (everything after the front-matter) into a Claude Cowork Project's custom instructions. Your steering rules become the Project's system prompt. The business logic is identical — only the delivery mechanism changes. Start using your skills in Cowork today while your tech team sets up the full Kiro automation stack.