· 8 min read
By the end of this guide, you will have a 4-agent system that ingests raw customer feedback from multiple sources, classifies it by sentiment and urgency, extracts recurring themes, identifies root causes, and produces a prioritized action report. The system handles support tickets, NPS responses, app store reviews, social media mentions, and survey results -- turning the noise of thousands of individual data points into a clear picture of what your customers need.
Most companies collect far more feedback than they can process. Support tickets pile up with recurring complaints that never get escalated. NPS comments sit in spreadsheets. App reviews contain feature requests that product teams never see. This agent team solves the processing bottleneck, ensuring that every piece of feedback is read, classified, and surfaced in a format that drives decisions.
You need Claude Code or the Claude Agent SDK installed and configured. For production use, the Agent SDK is recommended since you will likely want to run this analysis on a schedule (daily or weekly) as new feedback accumulates.
Prepare your feedback data in a structured format. The system works best when each feedback item includes: the raw text, the source channel (support ticket, NPS survey, app review, etc.), a timestamp, and any available customer metadata (plan tier, account age, segment). If your data is spread across multiple systems, consolidate it into a single feed before running the analysis.
Mission: Receive the batch of customer feedback, distribute it to specialist agents for analysis, collect their outputs, cross-reference findings across agents, and produce the final prioritized action report.
The Coordinator manages the overall workflow and handles the most important analytical task: connecting findings across agents. When the Sentiment Classifier flags a spike in negative sentiment about billing and the Theme Extractor identifies "unexpected charges" as a top-5 theme, the Coordinator connects these into a single, actionable finding rather than reporting them as separate observations.
Prompt guidance: Tell the Coordinator the business context: what product or service the feedback relates to, what recent changes might be driving sentiment shifts (a price increase, a new feature launch, a service outage), and what the team's current priorities are. This context helps the Coordinator distinguish between known issues and new signals.
Mission: Analyze each piece of feedback for sentiment (positive, negative, neutral, mixed), emotional intensity (mild, moderate, strong), and urgency (requires immediate response, can wait for batch processing, informational only). Produce aggregate sentiment metrics and flag outliers.
This agent reads every feedback item and assigns structured labels. It does more than simple positive/negative classification -- it distinguishes between a customer who is mildly annoyed ("the loading time could be better") and one who is about to churn ("I've been dealing with this issue for three months and I'm evaluating alternatives"). It also detects mixed sentiment, where a customer praises one aspect while criticizing another.
Prompt guidance: Provide calibration examples. Show the agent 5-10 feedback items with your desired classifications so it understands your thresholds. What counts as "strong negative" in your context? For a B2B SaaS company, "We're considering canceling" is critical. For a consumer app, "This is frustrating" might be the equivalent signal. Your calibration examples set the standard.
Mission: Identify recurring themes, topics, and feature requests across the feedback corpus. Cluster related feedback items into themes, quantify each theme's frequency and sentiment distribution, and track theme emergence over time.
The Theme Extractor moves from individual data points to patterns. It groups "the export takes too long," "CSV export times out on large datasets," and "why is downloading reports so slow?" into a single theme: "Export/download performance." It counts how many feedback items map to each theme, what the sentiment distribution is within each theme, and whether the theme is growing or stable compared to previous periods.
Prompt guidance: Specify the granularity you want. Too broad and every theme is "product quality" or "customer service." Too narrow and you get hundreds of micro-themes that are hard to act on. A good target is 10-20 themes for a batch of 500+ feedback items. Instruct the agent to create a hierarchy: top-level categories (Performance, Pricing, Features, Support) with specific themes nested underneath.
Mission: For the top themes identified by the Theme Extractor, investigate the underlying causes by analyzing the language patterns, customer segments affected, timeline of emergence, and correlation with product or business changes.
The Root Cause Analyst goes beyond "what are customers saying" to "why are they saying it." It examines whether a theme is concentrated in a specific customer segment (enterprise customers complaining about a feature that works fine for small teams), whether it emerged suddenly (correlating with a specific release or change), and whether the complaints describe a fundamental product gap or a usability issue with an existing feature.
Prompt guidance: Give this agent access to a product changelog or timeline of recent changes. If you shipped a new pricing model on March 1 and billing complaints spiked on March 5, the root cause is likely the pricing change, not a random fluctuation. Without this context, the agent can only speculate.
Customer feedback analysis benefits from a hybrid parallel/sequential flow:
The Sentiment Classifier and Theme Extractor run in parallel because they analyze the data from different angles and do not depend on each other. The Root Cause Analyst runs after both because it needs theme clusters (from the Extractor) enriched with sentiment data (from the Classifier).
Feedback analysis prompts need to handle messy, unstructured input. Customers do not write in clean sentences. They use slang, misspellings, sarcasm, and context that only makes sense if you know the product.
For the Sentiment Classifier: "You are a customer sentiment analyst. You will receive individual pieces of customer feedback. For each, provide: (1) sentiment: positive, negative, neutral, or mixed; (2) intensity: 1-5 scale where 1 is mild and 5 is extreme; (3) urgency: immediate, standard, or informational; (4) churn risk: high, medium, low, or none; (5) a one-sentence summary of the core message. Be alert to sarcasm -- 'Great, another update that breaks everything' is negative despite the word 'great.'"
For the Theme Extractor: "You are a customer feedback analyst specializing in theme identification. You will receive a batch of customer feedback items. Identify 10-20 recurring themes. For each theme, provide: (1) theme name (concise, descriptive), (2) parent category, (3) number of feedback items in this cluster, (4) representative quotes (3-5 per theme), (5) sentiment breakdown within the theme, (6) whether this theme is new, growing, stable, or declining compared to prior expectations."
For the Root Cause Analyst: "You are a product analyst investigating why customers are experiencing specific issues. For each theme you receive, analyze the clustered feedback to determine: (1) the most likely root cause, (2) which customer segments are most affected, (3) when the issue first appeared or intensified, (4) whether this is a product gap, a usability problem, a communication failure, or an expectation mismatch, (5) the estimated business impact in terms of churn risk and expansion friction."
The final deliverable is a prioritized action report. Structure it for decision-makers who need to allocate resources:
For a SaaS company processing 1,200 feedback items from the past month:
Executive Summary: Overall sentiment: 62% positive, 24% negative, 14% neutral. Negative sentiment increased 6 points from last month, driven primarily by two themes: onboarding complexity (new) and API rate limiting (growing). Three accounts flagged as high churn risk require immediate outreach.
Top Theme: Onboarding Complexity (87 mentions, 78% negative)
Emerging Signal: API Rate Limiting (23 mentions, 91% negative, new theme)
Real-time urgency routing. Add a fast-path for the Sentiment Classifier to immediately flag feedback that mentions legal action, data breaches, safety concerns, or explicit churn intent. These items bypass the normal batch workflow and go directly to the appropriate team.
Competitive intelligence extraction. Train the Theme Extractor to identify mentions of competitors. "We switched from [Competitor] because..." and "I wish you had what [Competitor] offers" are valuable competitive signals hiding in your feedback data.
Segment-specific analysis. Run separate analysis passes for different customer segments (enterprise vs. SMB, new vs. mature accounts, high-value vs. standard). The same product issue often manifests differently across segments, and the appropriate response differs too.
Longitudinal tracking. Store theme data over time and add a Trend Analyst agent that compares the current period to previous periods. This turns a point-in-time snapshot into a continuous monitoring system that detects when issues are getting better or worse.