A chatbot can write a reply about a leave policy. An agent can also check the balance, apply the rule and prepare the approval for you to sign off.
An HR officer asks for a childcare-leave approval note. Watch where a chatbot on its own stops, then switch to an agent.
Each task alone is easy; the chain is not. A chatbot can summarise a policy, classify a request and draft an approval. Ask for all three in one go on a real case and mistakes pile up, because input size, task complexity and output scope all grow together.
| Limit of a chatbot on its own | What it means at work |
|---|---|
| No live data | It cannot check a current leave balance, this month's budget actuals or a case's real status. |
| No actions | It cannot create records, route approvals, update HR or Finance systems or send an Outlook note. |
| No memory between chats | Every new conversation starts fresh. |
| Made-up details | It can invent clause numbers, entitlement figures or circular references when unsure. |
| Varying answers | The same question can get different answers, which is risky for compliance wording. |
Tool use is what lets an agent reach real systems. Follow one question about a claim from start to finish.
Function calling. The model does not run anything itself. It writes a structured request, such as which tool and which claim number. Separate software carries it out and hands the result back, so the answer comes from live data and not from the model's memory.
In Amazon Quick these tools are called actions. An administrator decides which ones are switched on. A standard way for agents to discover and call tools across systems, the Model Context Protocol (MCP), is covered in the optional Deep Dive.
Agents did not need a smarter model. They needed more system around it. Tap an idea, or a Quick building block, to see how it fits.
In Quick you build an agent from
| Idea | Without it | With it |
|---|---|---|
| Tool use | Can only chat about a claim | Checks the claim's status and drafts the notice to the officer |
| Reasoning | Ask for a compliance report, get noise | Plans: gather data, classify, look up policy, draft the update, ask for approval |
| Reading documents | The officer retypes a scanned form | Reads the form and cross-checks the guideline |
| Framework | Every team rebuilds the plumbing | Quick supplies it, with no code |
Beyond one agent. Quick Flows run multi-step automation on a schedule, an inbound email or a Quick event. Quick Research plans its own investigation across several sources. Quick Automate adds approval gates and versioning, is Enterprise tier, and your instructor will demo it rather than have you build it.
The engine. Amazon Bedrock provides the foundation models underneath Quick. You never touch it directly.
An agent is not a different model. It is the same model with more wrapped around it. Climb the steps and watch the abilities pile up.
| Ability | LLM | Assistant | Agent | Agentic system |
|---|---|---|---|---|
| Generate text | ||||
| Remember the conversation | ||||
| Answer from your documents | ||||
| Call tools and other systems | Limited | |||
| Plan multi-step work | ||||
| Take actions | Some | |||
| Coordinate with other agents |
Where this fits. The pre-read covered how AI reads and writes text (steps 1 and 2). This workshop moves into steps 3 and 4. You build an agent in F3 and design one for your own workflow in F4.
There is no autonomy setting in Amazon Quick. The level is whatever the agent's instructions say. Pick a level and watch the decisions change hands.
Often mistaken for more autonomy
| Level | Human role | Main way it fails | Safer for |
|---|---|---|---|
| L1 | Operator, sets the rules | Bad extraction, bad data downstream | Data entry, form digitisation |
| L2 | Reviewer, handles escalations | Wrong routing, delayed action | Triage, classification, routing |
| L3 | Supervisor, approves the output | Wrong plan, wasted effort | Reports, analyses, drafts needing sign-off |
| L4 | Strategic partner | Wrong goal, repeated at scale | Research, exploration, ideas |
You already have a name for these sentences. In F2 they are the T · Trust slot of CRAFT: the citation rules and the hard “do not” lines. Trust is where the level lives. In F4 you write it for a workflow of your own.
Why L3 and above needs care. Once an agent plans its own steps, new problems appear: looping, juggling several tools and recovering from errors. That is why the human checkpoint matters as much as the automation.
Eight admin workflows across HR, Finance, Procurement and Governance. Tap the one closest to your team.
| If your problem is… | Look at… | Start at |
|---|---|---|
| Repetitive data extraction from forms | Tool use and a fixed output format | L1 |
| Sorting inbound items by content | Classification and branching | L2 |
| Multi-step work with a clear goal | Step-by-step reasoning and tools | L3 |
| A scanned form plus document reasoning | Reading documents and tools | L3 |
| Strategic research and horizon scanning | A research-style agent | L4 |