Custom Agentic AI Foundations — 1-Day Workshop (All Staff / Track 1)

Learning Outcomes

By the end, participants can:

1. Explain in plain language what an LLM is and how it reads (tokenization), represents meaning (embeddings), and generates text (inference).
2. Write effective, structured prompts using the 4 pillars + persona.
3. Describe what makes something an agent and identify which workflow pattern fits a task they own.
4. Explain, in concept, how connectors / MCP let an agent reach documents, email, and systems.
5. Describe a good skill / system prompt and its guardrails.
6. Build a working assistant on AgentSea for the vendor-proposal workflow, including a verification step.

Agenda

Timings are adjustable; the capstone can grow to ~75 min by trimming M4/M6 if a deeper build is wanted. Full module descriptions follow below.

Module Detail

M0 · Welcome + meet the use case (15 min)

• Goals, agenda, AgentSea login, quick poll.
• Introduce the running scenario: “Your team receives vendor proposals every week. By 5pm today you’ll have built an AI assistant that reads one, checks it, and drafts a recommendation.” Everything that follows ladders up to this.

M1 · GenAI & the AWS AI landscape (30 min) — LLM basics

• What generative AI is, in plain language; foundation models; what an LLM can and can’t do.
• The AWS AI stack and where AgentSea sits: AgentSea (the app you use) → Amazon Bedrock (the engine) → Bedrock AgentCore (the agent runtime). A simple 3-layer “who works where” map.

M2 · How AI reads — Tokenization (30 min) — tokenization

• Characters → tokens → IDs; interactive BPE playground using vendor-proposal text (“SLA compliance”, “indemnification”).
• Why wording matters: same meaning, different token counts/behaviour. Light cost intuition (no Bedrock pricing deep-dive).

M3 · How AI captures meaning — Embeddings (30 min) — embeddings

• Tokens → vectors where similar meaning sits close; 3D embedding-space explorer.
• Use-case tie-in: how an assistant finds the relevant clause in a 40-page proposal even when the wording differs — the idea behind semantic search and knowledge spaces.

M4 · How AI thinks & generates — Inferencing (30 min) — inferencing

• Self-attention (what the model “looks at”), next-token prediction, and temperature = predictable vs creative output (when to use which for a compliance summary vs brainstorming).
• Framed simply: inference = how the model produces each next word.

M5 · Prompt Engineering + hands-on (75 min) — prompt engineering

• 4 pillars (Clear · Specific · Contextual · Constrained); persona technique; structured output (tables/checklists); negative constraints; brief Chain-of-Thought.
• Hands-on (on AgentSea): transform a vague “summarise this proposal” into a powerful prompt that returns a structured vendor summary + risk flags — a guided exercise on the vendor-proposal scenario.

M6 · From LLMs to Agents + Workflow Patterns (50 min) — agentic AI workflow

• Chatbot vs agent; the agent loop (Observe → Plan → Act → Reflect); levels of autonomy.
• 4 workflow patterns (Chaining · Parallelisation · Routing · Orchestration) with healthcare-admin scenarios.
• Map both use cases: Vendor Proposal = chaining (read → extract → validate → draft); Policy & Circular Compliance = routing + chaining (classify the circular → route to the affected policy/SOP owners → draft the updates).

M7 · Skills & System-Prompt Design on AgentSea (40 min) — skills

• AgentSea’s 3 skills today (Doc Analysis, Doc Generation, Outlook) + Document Builder.
• What makes a good skill / system prompt: role, instructions, output spec, guardrails, what NOT to do; when to formalise vs one-shot.
• Concept demonstration (subject to AgentSea platform readiness): custom, reusable skills that teams can create and share. Trainers may illustrate this as a concept — using another AI tool if it isn’t yet enabled in AgentSea.

M8 · Connectors & MCP — how agents reach your data (30 min) — connectors/MCP concept

• Why an agent needs to connect to documents, email, and systems; the MCP idea as the open “universal adapter” between AI and tools/data — taught tool-agnostically.
• In AgentSea: the Outlook connector today, with document-library (e.g. SharePoint), knowledge spaces, and multi-step workflow connectors shown as concepts demonstrated subject to platform readiness — trainers may illustrate these using another AI tool if not yet enabled. “Your DNA/IT team installs + authorises; you use the result.”
• Secondary use case lands here: the Policy & Circular Compliance Assistant retrieving from a policy/SOP library — shown as a concept (subject to platform readiness) to make the connector story concrete.

Capstone · Build the Vendor Proposal Assistant (50 min)

On current AgentSea (single agent + Doc Analysis + Doc Generation + system prompt):

1. Write the system prompt (role + 5–7 rules: SGD default, PII redaction, escalation threshold, never-invent-data, output format) — starter template provided.
2. Upload the synthetic proposal pack; run Document Analysis → structured summary.
3. Iterate on the deliberate-error proposal (line items don’t sum to the stated total) — the verification / agent-loop teaching moment.
4. Generate the recommendation memo (Document Generation).

• Optional close: capture the next agent idea on a mini Agent Design Canvas.

Wrap-up & what’s next (10 min)

• Takeaways; a look at where AgentSea is heading — document-library and knowledge-space connectors, multi-step workflows, and shareable skills (demonstrated as concepts, subject to platform readiness); champion network; feedback.