Agentic AI System Design / Practice
Design it under a clock
12 scenarios drawn from real agentic products. Each one runs like an interview: you get the prompt and the starting assumptions, you design, and the constraints that break your first answer arrive one at a time. The strong-answer signals stay hidden until you finish, because reading them first turns practice into recognition.
Everything you type stays in this browser. There is no account, no upload, and no grading service. The scorecard and the rubric are how you grade yourself.
-
01
Cloud Security Investigation and Remediation Agent
authoritytool safetydurable executionincident response
-
02
Enterprise Coding Agent for a Million-Line Repository
repository contextsandboxinglong-running executionevaluation
-
03
Insurance Claims Intake and Adjudication Support Agent
documentshuman authoritylong-running workflowsgovernance
-
04
Customer Support Agent with Refund Authority
interactive latencyaction authorityscalefraud controls
-
05
Enterprise Research Agent with Document ACLs
retrievalaccess controlcitationsmemory
-
06
Long-Running Procurement and Vendor Coordination Agent
long-running workflowsexternal communicationapprovalsmulti-agent boundaries
-
07
Healthcare Prior-Authorization Preparation Agent
sensitive dataevidence and policyurgent workflowshuman authority
-
08
Production Incident-Response Copilot
real-time operationsreversible actionsobservabilityhuman coordination
-
09
Financial Research and Trade-Proposal Agent
freshnessevidencemarket-data licensingaction separation
-
10
Autonomous Data Analyst over an Enterprise Warehouse
code executiondata governancesemantic layercost control
-
11
Travel Disruption Recovery Agent
burst trafficexternal transactionsuser preferencesmulti-step compensation
-
12
Enterprise Agent Platform for Multiple Product Teams
platform boundariesmulti-tenancygovernancedeveloper experience
How to use these
- Read the prompt and the assumptions, then start the timer. Do not scroll ahead. The point is to design without knowing which constraint lands next.
- Clarify before you draw. Five minutes of requirements and risk beats twenty minutes of boxes. Name the risk tier and what the model may never decide.
- Open one reveal at a time. Update the design after each before you look at the next, the way the constraint would actually land in the room.
- Finish, then reveal. Score yourself against the required checks first, then read the strong-answer signals and the model outline. The gap between them is your study list.
If a scenario exposes a concept you cannot explain, the concept coverage matrix maps every idea in the course back to the module that teaches it.