Starts this lesson and continues through 11 more to the end of the course.

Flows, Automate, and agentic development

Why this matters

This is the "agentic" half of the suite — the part where Quick stops answering questions and starts doing things: filling web forms, calling APIs, writing to Salesforce, escalating to a human when stuck. Two products cover it, and AWS's own comparison is the cleanest summary you'll find (Automations overview, verified 2026-08-26):

"Amazon Quick Flows empowers any business user to automate repetitive, routine tasks without needing technical expertise. Amazon Quick Automate enables power users to orchestrate end-to-end business processes at enterprise scale."

Flows are "personal and team productivity tasks, run by individual users"; Automate is "centralized, enterprise-wide automations, especially those which need to run independent of an individual user," with "control flow logic, case management, error handling, custom code, and more." Learn the boundary and you'll route every request to the right tool in one sentence.

Quick Flows

Four ways to create one (creating-flows, verified 2026-08-26): natural language ("describe your desired flow in plain language, and Amazon Quick Flows generates a flow"), the visual editor, duplicating an existing flow, or from a chat agent conversation — a chat that worked becomes a repeatable flow.

Step types (same page):

User input steps (text/files)
AI response steps — "generate content from chat agents, web search, general knowledge, or image generation"
Data insight steps — "responses from Quick data or dashboards and topics"
Reasoning groups — "conditions, loops, or validation"
Action steps — "read or write operations in connected applications" (action-steps, verified 2026-08-26). Parameters come from the prompt, from @ references to previous steps, or are "acquired dynamically during flow execution." ⚠️ List operations return only the first batch — "Quick retrieves only the first batch to keep your flow fast and responsive." A flow is not an ETL tool; don't let anyone paginate a CRM through one.

Validity: unique name ≤128 chars, and "a valid flow must have at least one output step or action step." Sharing: Viewer ("can run the flow… only see the published version") vs Co-owner (edits, publishes, sees drafts; "only one person can edit at a time"); org settings can require an approval review before sharing (sharing-flows, verified 2026-08-26).

Where flows run: run mode, directly from chat, and — the sleeper feature — the browser extension, which can "run pre-defined workflows directly within your web browsing context" and "pass web page content as input" (flows-in-browser-extension, verified 2026-08-26).

The numbers that matter, from the dedicated limits page (verified 2026-08-26):

Limit	Value
Flows per instance / per user	10,000 / 100
Steps per flow	35
Iterations per reasoning group	50
Context limit	"Up to 1M characters"
Schedules per user / per instance	20 / 10,000

⚠️ Schedules (and image generation) are Region-limited: us-east-1, us-west-2, eu-west-1 only (same page). If your home Region is elsewhere, scheduled flows are not a thing you can promise.

Quick Automate

Automate is "an AI-powered application that creates sophisticated automations using natural language or documentation… by uploading process documents or simply describing their business process" (using-amazon-quick-automate, verified 2026-08-26). Structure: automation groups (permissions/resources) → projects ("organizing, authoring, testing, deploying, and maintaining automations") → automations on a Visual Designer canvas with an Actions panel "organized by category (Agents, Process flow, Web browser, etc.)" (building-automations, verified 2026-08-26).

The two agents are the headline:

UI agent — "a native agent that understands natural language instructions to perform complex browser actions. It can autonomously navigate websites, click, type, read data, and produce structured outputs." This is computer-use-style browser automation as a managed AWS feature — the part of Quick with no QuickSight ancestor at all.
Custom agent — tool-calling over integrations: "integrations as its tool interface… Code as tool, and other native actions like human-in-the-loop task." Documented tools (automate-custom-agents, verified 2026-08-26): Amazon S3, Bedrock Data Automation, Comprehend, Textract, Custom REST API, Custom MCP connector, Microsoft Outlook, Salesforce, plus Code and HITL tasks.

Model choice, same page: managed modes Fast ("simple tasks like summarization, classification") and Pro ("complex tasks that involve reasoning and orchestration of multiple tools") — underlying models unnamed — or Custom mode: bring your own Bedrock model; the docs' example is us.anthropic.claude-3-5-sonnet-20241022-v2:0, with "model inference… billed separately to the account associated with that Bedrock integration."

Engineering discipline is built in (deploying-automations, verified 2026-08-26):

Typed variables (String/Number/Boolean/Array/Object/File/Data table); input/output schemas validated "before execution starts"; file inputs max 5 MB.
Version control: Live → Committed → Deployed; "only committed versions can be deployed."
Human-in-the-loop: at deployment you "select the resolver users or user groups for each of the tasks that require human-in-the-loop"; a stuck custom agent "trigger[s] a Human-in-the-Loop (HITL) task… visible in the task center."
Credentials are first-class: website credentials for UI-agent logins and action credentials for integrations, "securely stored and encrypted… without exposing them to authors or end users."
Triggers: schedules (cron supported, start within 15 min) capped at "10 parallel executions per trigger and 50 across all automations within an account" — and API triggers:

aws quicksight start-automation-job \
  --aws-account-id 111122223333 \
  --automation-group-id <group> --automation-id <automation> \
  --input-payload '<json>'          # async; returns JobId
aws quicksight describe-automation-job ...   # QUEUED|RUNNING|SUCCEEDED|FAILED|STOPPED

StartAutomationJob is POST /accounts/{AwsAccountId}/automation-groups/{id}/automations/{id}/jobs, payload up to 7,000,000 characters; IAM actions quicksight:StartAutomationJob / DescribeAutomationJob on ARNs like arn:aws:quicksight:…:automation-group/GROUP/automation/AUTOMATION (verified 2026-08-26). Flows have API surface too — GetFlowMetadata (PublishState PUBLISHED | DRAFT | PENDING_APPROVAL) is verified, and list-flows / search-flows / update-flow-permissions exist in the CLI; we did not verify the full flow-authoring operation set, so check the API reference before promising CreateFlow-style IaC.

⚠️ Gates: Automate requires an account "provisioned through the AWS Management Console" and is Professional/Enterprise only (using-amazon-quick-automate + pricing, verified 2026-08-26). And both products burn agent hours (lesson 1) — an Automate process that runs hourly is a meter, not a feature.

Choosing, in one breath

Expense-report chaser for yourself or your team → Flow. Invoice-processing pipeline with approvals, retries, credentialed website access, and an API trigger from your ERP → Automate. The tell-tales: "runs as a person" vs "runs as the org", 35 steps vs projects-and-versions, shared flow vs deployed automation.

Check yourself

A flow needs to process every open Jira ticket. What documented behavior breaks this, and what's the right tool?
What can Automate's UI agent do that nothing in Flows can?
Where do the credentials for a UI-agent website login live, and who can read them?
An automation edit isn't taking effect in production. Walk the version-control explanation.
Your home Region is ap-southeast-2 and a user wants a daily scheduled flow. Problem?
Name the two managed agent modes in Automate's custom agents, and what Custom mode changes about billing.

Answers

Action list-steps "retrieve only the first batch" — a flow won't paginate. Batch/complete processing belongs in Automate (or a real pipeline).
Autonomously drive a web UI — navigate, click, type, read data into structured outputs. Flows' browser extension only runs flows with page content as input; it doesn't drive the browser.
In Automate's credential store ("website credentials"), encrypted, and "without exposing them to authors or end users."
Live (editable) → must be committed → only committed versions can be deployed. An uncommitted or undeployed edit never reaches production.
Yes — flow schedules are documented as available only in us-east-1, us-west-2, and eu-west-1.
Fast and Pro (models unnamed in docs). Custom mode brings your own Bedrock model, and inference bills separately to the account behind that Bedrock integration.

Teaching this section

Open with the Flows-vs-Automate quote and make every exercise start by classifying a task.
Demo: build a two-step flow from natural language, run it from chat; then show (video or console) a UI agent filling a form — the room changes when the browser moves by itself.
Good discussion: "What's the first process in your org you'd trust to a UI agent, and what HITL gate would you demand?"
Watch for: learners equating Flows' browser extension with the UI agent — one runs flows in a page context, the other autonomously operates the page.
Timing: 30 min presented; the classify-the-task drill is worth 10 on its own.

← PreviousTopics, dashboard Q&A, and natural-language BI Next →Research, auditability, and running the meter