Real-world agents are defined by three things: the tools they reach for, the memory they carry, and the reasoning steps they follow. In this lab you deconstruct a live use case โ a Customer Insight Analyzer โ from the inside out.
Learning Objectives
01
Read a real-world agent use case and identify its inputs, goals, constraints, and success criteria.
02
Classify which tools the agent uses โ data connectors, reasoning tools, and delivery mechanisms.
03
Map which type of memory โ short-term, long-term, or working โ each information need requires.
04
Sequence the agent's full reasoning flow from data ingestion through validated delivery.
Your Lab Roadmap
1
Case Brief
Read the Customer Insight Analyzer scenario โ explore context, goal, constraints, and data sources
2
Tool Kit Mapping
Drag 12 tools into the correct layer โ Data In ยท Reasoning ยท Action ยท Delivery
3
Memory Classification
Assign 9 information needs to Short-term ยท Long-term ยท Working memory
4
Reasoning Sequencer
Drag 7 reasoning steps into the correct execution order
โ
Agent Canvas
4 synthesis questions โ consolidate your full technical deconstruction
5
Knowledge Check
7 questions ยท agent anatomy ยท tools, memory, and reasoning ยท 70% to pass
โ
Your Results
Score breakdown ยท key takeaways
Section 01 ยท Read Before You Build
The Case Brief
Before deconstructing any agent, you must understand the problem it solves. Explore all four context panels โ each informs a different part of your technical analysis.
๐ Scenario
๐ฏ Goal
๐๏ธ Data Sources
โ ๏ธ Constraints
The Customer Insight Analyzer
Acme SaaS has 50,000 customers and a support team drowning in tickets. Every quarter, the Product team spends three weeks manually reading Zendesk tickets, App Store reviews, and Salesforce CRM notes to produce a Customer Insight Report โ a document that shapes the next quarter's roadmap.
The process is slow, inconsistent, and expensive. Important signals get missed. The Product team suspects there are patterns in the noise โ recurring complaints, emerging feature requests, churn predictors โ that no human has the bandwidth to surface at scale.
They want an autonomous agent that ingests all three data sources continuously, identifies patterns, generates strategic insights, and delivers a structured report to the product leadership team every Monday morning โ with no manual intervention.
Primary Goal
Produce a weekly Customer Insight Report containing the top 5 emerging themes, top 3 churn risk signals, and the top 2 feature requests โ ranked by frequency and sentiment intensity โ without any human labor in the loop.
Success Criteria
โ Report delivered every Monday before 8:00 AM
โ Each insight is backed by a minimum of 3 source citations
โ Sentiment scoring is consistent week-over-week
โ No PII (personally identifiable information) included in the final report
โ Product team spends less than 10 minutes reviewing โ not rewriting
Why an Agent (not a script)?
A script could pull data. An agent can reason about it โ classifying, comparing to prior weeks, surfacing non-obvious connections, and adjusting its analysis when data quality is poor.
Data Source 1 โ Zendesk (Support Tickets)
~2,000 new tickets per week. Plain text, tagged by category. Contains customer language, frustration signals, and unstructured feature requests. Requires API authentication.
Data Source 2 โ App Store Reviews
~300 new reviews per week across iOS and Android. Public data โ no auth required. Contains star ratings + free-text commentary. High signal-to-noise ratio for product quality perception.
Data Source 3 โ Salesforce CRM Notes
Customer success rep call notes from ~150 accounts per week. Highly valuable but inconsistently formatted โ some reps write paragraphs, others write bullet points or nothing at all. Requires Salesforce API connector with OAuth.
Total weekly input volume: ~2,450 documents ยท ~1.2M tokens before chunking.
Privacy Constraint
No customer names, email addresses, or account IDs may appear in the final report. The agent must strip or anonymize PII at the data ingestion stage.
Reliability Constraint
If fewer than 500 tickets are retrieved in a week (indicating an API failure), the agent must not generate a report. It should flag the error and notify the ops team.
Cost Constraint
The total LLM cost per weekly run must stay under $4.00. This forces the agent to use chunking and summarization strategies rather than passing raw full-text to the LLM.
Consistency Constraint
The agent must compare this week's insights against the prior 4 weeks to distinguish "new signal" from "persistent issue." This requires persistent memory across runs.
๐ Before moving on: Make sure you've read all four panels. The constraints panel in particular directly shapes which tools, memory types, and reasoning steps the agent needs.
Section 02 ยท Lab Activity
Tool Kit Mapping
Every agent capability requires a tool. Drag each tool into the layer where it operates โ Data In, Reasoning, Action, or Delivery. Use the Case Brief if you need to re-read the scenario.
Unassigned tools โ drag to the correct layer
On mobile? Tap to assign
Section 03 ยท Lab Activity
Memory Classification
The right memory architecture prevents the agent from re-doing work, losing context, or forgetting historical patterns. Classify each information need into the memory type that handles it best.
โก
Short-Term
Current context window โ what's happening right now in this run
๐๏ธ
Long-Term
Persists across runs โ vector store, database, or retrieval system
๐
Working
Task state โ tracks progress, completed steps, and pending actions
Unassigned information needs โ drag to the correct memory type
On mobile? Tap to assign
Section 04 ยท Lab Activity
Reasoning Sequencer
A well-designed agent executes steps in a deliberate order. Drag the reasoning steps below into the correct sequence โ from first action to final delivery.
๐ก Hint: Think about what the agent must know before it can do each subsequent step. What must be true before you can classify? Before you can compare? Before you can deliver?
Synthesis Activity
Agent Canvas
Consolidate your full technical deconstruction. For each design decision, choose the best answer based on what you learned in the Case Brief and previous sections.
Section 05
Knowledge Check
Test your understanding. You need 70% or higher to mark this section complete.