← Back to Workshop Hub FOUNDATIONS · MODULE F1

From Chatbots to Agents

A chatbot can draft a reply about a leave policy — but it can't check the staff member's balance, apply the rule, update the record, or notify payroll. Discover the handful of ideas that turn a language model into an agent that takes work off your plate.

📖 ~12 min read ⚡ Interactive 🏥 Healthcare Admin L100 Foundational

🚧 Great at Words, Stuck at Work

A large language model — the engine behind any chatbot — is brilliant at generating text. But on its own it works in isolation: no memory between conversations, no access to your live systems, no ability to take an action. For a busy healthcare organisation, that gap is the difference between an answer and an outcome.

🏥 The Leave Request Problem
An HR officer asks the assistant: "A staff member applied for 12 days of childcare leave. Does our guideline allow that, what's their remaining balance, and can you draft the approval note?" A plain chatbot will write a confident, sensible-sounding reply. But behind the curtain, here's what it cannot actually do:
// Leave request received for staff EMP-2218...
look_up_leave_record("EMP-2218")❌ NO SYSTEM ACCESS
check_leave_balance("childcare")❌ NO LIVE DATA
read_policy_rule("childcare-eligibility")❌ NOT IN TRAINING DATA
create_approval_record("EMP-2218")❌ NO ACTION CAPABILITY
notify_payroll("primary_on_call")❌ NO MESSAGING ACCESS
// The chatbot can only reply: "You should check the leave record and notify payroll..."

The chatbot produced words. The officer needed a completed workflow.

🧭 Where the Workflow Breaks

Click any step to see exactly where a standalone chatbot hits its wall — and what an agent adds at each stage.

ONE LEAVE REQUEST · FIVE STEPS 📥 Request arrives 🔎 Look up the record 📋 ⚠️ Apply the rule ⚙️ Update the system 📣 Notify payroll

⚠️ The Fallacy of Composition

A chatbot can summarise a policy. It can classify a request. It can draft an approval. So surely it can do all three in sequence on a real case? Not reliably. Each task fits inside what the model can do alone — but chain them and reliability drops sharply as input size, task complexity, and output scope grow together.

~85%
accuracy on a single task
~60%
on 3 chained tasks
~30%
on 5+ step workflows
3–5×
cost multiplier from retries

Illustrative — the exact numbers vary by model and task, but the shape is always the same.

📏

Input Size

Long policy manuals, multi-department circulars and full procurement packs exceed the context window or degrade quality.

🧩

Task Complexity

Multi-step reasoning (look up record → check rule → assess → draft → notify) overwhelms a single chatbot pass.

📐

Output Scope

Producing an approval + updated record + payroll alert + audit entry in one shot multiplies the chance of errors.

🫙 Other Limits of a Chatbot on Its Own

No live dataCan't check a current leave balance, this month's budget actuals, or a case's real status.
No actionsCan't create records, route approvals, update HR/Finance systems, or send an Outlook notification.
No persistent memoryForgets prior context between sessions; treats every conversation as fresh.
HallucinationsInvents clause numbers, entitlement figures, or circular references when it's unsure.
Non-deterministicThe same question can produce different answers — risky for compliance or contractual outputs.
💡
The mental model: a chatbot is a brain in a jar — it can think and talk, but it can't reach into your policy library, your HR system, or your inbox. Agents give the brain hands and eyes.

🔑 Four Ideas Turn a Chatbot into an Agent

Agents didn't need a smarter model — they needed a smarter system wrapped around the model. Four ideas make the difference. The one that starts it all is tool use: giving the model a way to reach real software. Watch how that works, then explore all four.

HOW AN AGENT USES A TOOL — "What's the status of claim CLM-22-184?" 🧑 You ask plain question 🧩 Agent wraps it offers tools 🔌 Model calls tool get_claim(id) 🗄️ System replies real data 💬 Grounded answer from live data
🔧
The staff member sees "the AI looked up the claim." Under the hood, ordinary software did the lookup and the model directed it — this is function calling, and Amazon Bedrock's Converse API supports it natively across Claude, Nova, Llama and Mistral. The emerging Model Context Protocol (MCP) is how agents discover and call tools across systems (covered in the Deep Dive).

🧠 All Four Ideas

Click any card to see how it fixes a limit you met in Tab 1.

🔌

Tool Use (Skills)

The agent can call your HR system, query a claim, or post an Outlook note. On AgentSea, these are its skills.

🧠

Step-by-Step Reasoning

Break a goal into sub-tasks: "process this request" becomes look up → check → draft → notify.

👁️

Reading Docs & Images

Reason across text, scanned forms and tables together — one model reads the guideline PDF and the form.

🏗️

Agent Frameworks

The scaffolding — memory, planning loops, tool orchestration — so every team doesn't rebuild it.

🧰 On AgentSea, Tool Use = Skills

You don't wire up function calling yourself. On AgentSea, tool use shows up as ready-made skills you switch on for your agent. Today's live set:

📄 Document Analysis 📝 Document Generation / Document Builder 📧 Outlook 🔗 Knowledge spaces & connectors CONCEPT — subject to platform readiness
🏗️
AgentSea is built on Amazon Bedrock AgentCore — the managed agent runtime that handles memory, planning and observability for you. The open-source Strands Agents SDK is the same idea for teams that build their own. You'll pick a skill and give your agent its instructions in F3.

🗺️ Each Idea, Mapped to Real Admin Work

IdeaWithout itWith it — a healthcare-admin example
🔌 Tool UseCan only chat about a claimReads the claims system, checks status, updates the record, notifies the officer via Outlook
🧠 ReasoningAsk for "a compliance report", get noisePlans: gather data → classify → look up policy → draft update → request approval
👁️ MultimodalOfficer must retype a scanned formReads the form, extracts the fields, cross-references the guideline
🏗️ FrameworksEvery team rebuilds tool-callingBedrock AgentCore / Strands handle memory, retries, observability, guardrails
Key insight: none of these make the model "smarter" in a vacuum. They make the system smarter — so it can perceive, reason step-by-step, and act. The intelligence is in the loop, not just the model.

📈 Four Steps from Answering to Acting

"Agent" isn't a different model — it's a different amount of system wrapped around the same model. Each step adds capability. The same Claude or Nova model can power any of them. Click a step to see what it can and can't do.

SAME MODEL UNDERNEATH · MORE SYSTEM AROUND IT → 🧠 LLM brain in a jar 💬 Assistant + memory + docs 🤖 Agent + tools + planning 🕸️ Agentic System + coordination

📊 Side by Side

CapabilityLLMAssistantAgentAgentic System
Generate text
Remember conversation context
Look up documents (RAG)
Call tools / external systemsLimited
Plan multi-step workflows
Take actionsSome
Coordinate with other agents
Self-monitor & evaluateSome
🎯
Stages 2–4 all use the same foundation model. What separates them is the architecture around it — retrieval, tools, planning, coordination. You don't need a bigger model to move up; you build a better system around it.
📚

The Pre-read

How AI Works covered how AI reads, represents meaning, and generates text — Stages 1–2.

🎯

This Workshop

We move into Stages 3–4 — agents that plan and act. You'll build one on AgentSea in F3.

🚀

Your Use Case

In F4 you design an agent for your own workflow on the Agent Design Canvas — usually Stage 3 to start.

🎚️ Autonomy Is a Dial, Not a Switch

Not every "agent" is equally independent. Some follow a fixed script; others set their own path to a goal. This four-level model helps you scope what's appropriate for each task — and decide where the human stays in the loop. Click a level for a healthcare-admin example.

MORE AUTONOMY → (AND MORE HUMAN OVERSIGHT TO DESIGN IN) L1 🔒 Rule-based fixed steps L2 🔀 Dynamic workflow routing & branching L3 🎯 Goal-driven plans its own steps L4 🌐 Full autonomy self-directed

Most production agents in healthcare admin today sit at L3 — goal-driven, with a human approving the output.

🧭 Picking the Right Level

Start at L1 or L2

Automate the boring, high-volume tasks first — form digitisation, triage. Build trust, then climb the ladder.

🎯

L3 for High Value

Reserve goal-driven planning for tasks where the goal is clear but the path varies — variance commentary, vendor analysis, circular gap-checks.

⚠️

L4 with Caution

For compliance-sensitive work, keep a human in the decision loop. Use L4 patterns for research and ideation, not autonomous action.

Autonomy vs. Risk — a Mental Model

LevelHuman roleMain failure modeSafe domains
L1Operator (defines rules)Bad extraction → bad data downstreamData entry, form digitisation
L2Reviewer (handles escalations)Wrong routing → delayed actionTriage, classification, routing
L3Supervisor (approves output)Wrong plan → wasted effortReports, analyses, drafts needing sign-off
L4Strategic partnerMisaligned goal → wrong direction at scaleResearch, exploration, ideation
🏗️
Why frameworks matter at L3+: once an agent plans its own steps you face new problems — runaway reasoning loops, juggling several tools, and error recovery. That's what Strands Agents and Bedrock AgentCore solve — the deeper technical stack covered in the Technical Community track.

🏥 Where Agents Land in Healthcare Admin

The theory becomes real in your day-to-day. Here are eight workflows across HR, Finance, Procurement and Governance that move from "the chatbot answers a question" to "the agent gets the work done." Tap the one closest to your team.

📋Guideline AssistantGovernance · All staff
💰Vendor Proposal AnalysisProcurement
📜Policy & Circular ComplianceGovernance
🔄Claims / Leave ProcessingHR
📊Budget Variance CommentaryFinance
📧Inbox TriageAll · Admin
📝Minutes → ActionsAll staff
📚Onboarding KnowledgeHR · L&D
📋
The Guideline Assistant is the one everyone builds together in F3 — it's universal, it teaches grounding, and it's the template for the agent you'll design on your own use case in F4.

🧩 A Simple Decision Guide

Match the shape of your problem to a starting pattern and autonomy level.

If your problem is…Look at…Start at
Repetitive data extraction from formsTool use + structured outputL1
Routing inbound items by contentClassification + branchingL2
Multi-step workflow with a clear goalStep-by-step reasoning + toolsL3
Cross-system report needing synthesisOrchestrator + worker stepsL3
Scanned form + document reasoningMultimodal + toolsL3
Strategic research & horizon-scanningResearch-style agentL4

🎁 Four Things to Take Away

🚧

A chatbot isn't enough

It generates text but can't perceive, act, or remember. Reliability drops fast across chained steps — the fallacy of composition.

🔑

Four ideas make an agent

Tool use (skills), step-by-step reasoning, reading docs & images, and an agent framework — the intelligence is in the loop.

📈

Capability is layered

LLM → Assistant → Agent → Agentic System. Same model underneath; smarter architecture around it.

🎚️

Autonomy is a dial

Match the level (L1–L4) to your risk. Most production admin agents live at L3, with a human approving the output.

💡
The leadership read
Chatbots make people faster. Agents take work off the plate. The shift from "help me write this" to "handle this for me, and I'll approve it" is where the real time savings live — and why designing the human checkpoint matters as much as the automation.

🚀 What's Next in the Workshop

You now know what an agent is. Next you'll learn to steer one, build one, and design your own.

F2 · NEXT

Prompting Your Agent

The 4 pillars + persona — writing the system prompt that steers your agent.

F3 · BUILD

Build on AgentSea

Give it a skill, ground it in a document, and produce a brand-compliant briefing.

F4 · DESIGN

Agent Design Canvas

Design an agent for your own workflow — the bridge into the optional Use Case Deep Dive.

🧭
Curious about the four workflow patterns (chaining, parallelisation, routing, orchestration) and how agents connect to your systems via MCP? Those go deeper in the optional Use Case Deep Dive.