Customer support agent mind maps
Everything the CCA-F exam asks about the customer support agent, on one page. It starts with which exam scenarios bring these questions. Then six mind maps, each in the form "what you see, where the fault is, the fix", with the rule to remember, the exam-guide line it comes from, and one practice question.
Domain 1 · 27%Domain 5 · 15%Guide 1.1, 1.4, 1.5, 5.1, 5.2About 30 minutes
Which scenarios ask about the support agent?
The exam shows you 4 scenarios out of a bank of 6. One is the support agent itself. The same ideas come up elsewhere:
| # | Scenario | Domains | Support-agent questions? | Maps to study |
|---|---|---|---|---|
| 1 | Customer Support Resolution Agent | D1, D2, D5 | Yes, the core one. Agent SDK, MCP tools get_customer, lookup_order, process_refund, escalate_to_human, target of 80%+ resolved on first contact | All six, plus MCP maps 3 and 4 |
| 3 | Multi-Agent Research System | D1, D2, D5 | Partly. The same agent loop and hooks run inside every subagent | 1, 2 |
| 2, 4, 5, 6 | Claude Code and extraction scenarios | — | Hooks vs instructions shows up in Claude Code too | 2 |
Guide §5, scenario 1: “Customer Support Resolution Agent — Agent SDK; custom MCP tools (get_customer, lookup_order, process_refund, escalate_to_human); target 80%+ first-contact resolution with escalation. (D1, D2, D5)”
First, the picture: the loop and where hooks sit
A prompt asks Claude to do something, and it usually does. A hook makes it happen every time, because it's code that runs outside Claude. Many questions in this scenario come down to “ask” or “make”.
The six maps at a glance
| # | Question | Usual answer | Practise |
|---|---|---|---|
| 1 | When does the loop stop? | On stop_reason: continue on tool_use, stop on end_turn | stop_reason |
| 2 | Prompt or hook? | Must happen every time, or money is at stake: a hook or a prerequisite gate | Prompt or hook |
| 3 | Tool results come in messy formats | A PostToolUse hook cleans them before Claude sees them | Hook cleanup |
| 4 | When should the agent escalate? | Explicit criteria and examples. Escalate at once when the customer asks for a human | Who escalates |
| 5 | Several customers match | Ask for another identifier. Never guess | Several matches |
| 6 | Long chats lose exact facts | A “case facts” block, trimmed tool results, key facts at the start | Context meter |
A yellow row in each map is the runner-up: the case where a different answer wins because one fact in the question changes.
Mind map 1: When does the loop stop?
| What you see | Where the fault is | Fix |
|---|---|---|
| The loop stops when Claude's text says “I'm done” | It reads the wording, not the signal | Check stop_reason |
| The loop stops after 5 turns even when work remains | A fixed turn limit is the main way of stopping | Stop on end_turn. A limit is only a safety net |
stop_reason is tool_use | — | Run the tool, add the result to the history, and call Claude again |
stop_reason is end_turn | — | Stop and return the answer |
| The same steps happen every time, in a fixed order | The model doesn't need to choose | A fixed sequence in code can beat a model-driven loop |
stop_reason is tool_use, and stop at end_turn. Never decide from the text, and never use a turn limit as the main way to stop.Guide 1.1: “Continue the loop when stop_reason == "tool_use"; terminate when "end_turn". Anti-patterns: parsing natural-language signals to determine termination; arbitrary iteration caps as the primary stop; checking assistant text content as a completion indicator.”
- Tell Claude to always end with “DONE” when finished
- Continue while
stop_reasonistool_use, and stop onend_turn - Always run at least 5 turns
- Check the reply for words like “refund processed”
Mind map 2: Prompt or hook?
| What you see | Where the fault is | Fix |
|---|---|---|
| In 12% of chats the agent looks up orders before checking the customer's identity | The order is only asked for in the prompt | A prerequisite gate: block lookup_order and process_refund until get_customer returns a verified ID |
| Refunds over $500 must never be processed automatically | A prompt rule fails sometimes | A hook on the outgoing call blocks it and routes to escalate_to_human |
| Someone suggests a stronger system prompt or few-shot examples for a must-never rule | Both still leave a small failure rate | Use code (a hook or gate) for rules that must always hold |
| It's about tone or style, like “be friendly and brief” | A hook can't enforce tone, and slips don't cost money | A prompt instruction is right here |
Guide 1.4: “When deterministic compliance is required (e.g., identity verification before financial ops), prompt instructions have a non-zero failure rate. Block downstream tool calls until prerequisites complete.” Guide 1.5: “Hooks give deterministic guarantees vs prompt-based probabilistic compliance. Block policy-violating actions (e.g., refunds > $500) and redirect to escalation.”
- Put the rule in capital letters at the top of the system prompt
- Add five few-shot examples of large refunds being escalated
- Add a hook that blocks
process_refundabove $500 and callsescalate_to_human - Remove
process_refundfrom the agent's tools
Mind map 3: Tool results come in messy formats
| What you see | Where the fault is | Fix |
|---|---|---|
| Dates come back as Unix timestamps from one system and ISO 8601 from another, and Claude mixes them up | Claude sees raw, mixed formats | A PostToolUse hook converts them to one format first |
| Status comes back as number codes (3, 7, 12) | Claude has to guess what they mean | The hook turns codes into words (“shipped”) |
| A tool returns 40 fields when 5 matter | Context fills up with noise | Trim the result to the fields that matter (the hook can do it) |
| The problem is in the outgoing call, not the result | A post-call hook runs too late | A hook on the outgoing call (map 2) |
PostToolUse hook, before Claude sees them. Asking Claude in the prompt to “convert all dates” is the weaker answer.Guide 1.5: “PostToolUse hooks intercept results for transformation before the model sees them. Normalize heterogeneous formats (Unix ts, ISO 8601, numeric codes).”
- Explain both formats in the system prompt
- Add a
PostToolUsehook that converts all dates to one format - Add a
convert_datetool for the agent to call - Ask the orders team to change their API
Mind map 4: When should the agent escalate?
| What you see | Where the fault is | Fix |
|---|---|---|
| The agent escalates easy cases and tries to handle hard policy exceptions itself | The line between “resolve” and “escalate” is unclear | Explicit escalation criteria plus few-shot examples of each |
| The customer says “let me talk to a person” | — | Escalate right away |
| The customer is frustrated, but the fix is simple | Escalating would be slower for them | Offer the fix first. Escalate if they still ask for a person |
| Escalation is routed on sentiment or on Claude's own confidence score | Neither measures how complex the case is | Route on explicit criteria: policy gaps, exceptions, no way forward |
| Humans get escalations and have to read the whole chat | No handoff summary | A structured summary: customer ID, root cause, amount, recommended action |
Guide 5.2: “Triggers: customer requests a human; policy exceptions/gaps; inability to progress. Escalate immediately on explicit demand; offer resolution when straightforward. Sentiment-based escalation and self-reported confidence are unreliable proxies for complexity.” Guide 1.4: “compile structured handoff summaries (customer ID, root cause, refund amount, recommended action).”
- Escalate when the customer's sentiment turns negative
- Escalate when Claude rates its confidence below 70%
- Write explicit escalation criteria, with examples of cases to resolve and cases to escalate
- Train a separate classifier on past tickets
Mind map 5: Several customers match
| What you see | Where the fault is | Fix |
|---|---|---|
get_customer("Priya Sharma") returns three accounts | The agent can't know which one | Ask the customer for another identifier (email, order number) |
| The agent picks the most recently active account | A guess that will sometimes be wrong | Never choose by a heuristic. Ask |
| Exactly one account matches | Nothing is ambiguous | Go ahead with that account |
Guide 5.2: “Multiple customer matches → ask for identifiers (not heuristic selection).”
get_customer returns two accounts in the same city. What should the agent do?
- Use the account with the most recent order
- Ask the customer for their email or an order number
- Escalate to a human
- Show both accounts' details to the customer and ask which one
Mind map 6: Long chats lose exact facts
| What you see | Where the fault is | Fix |
|---|---|---|
| After summarizing, “$247.50 refund promised by 5 March” becomes “a refund was discussed” | Summaries drop numbers, dates and promises | Keep a “case facts” block with exact values, outside the summary |
| Details in the middle of a long context get missed | “Lost in the middle” | Put key facts at the start, with clear section headings |
| Each order lookup adds 40 fields and the context fills up | Tool results are kept in full | Trim tool results to the fields that matter |
| The chat is short and nothing has been summarized | No facts have been lost | No extra structure needed yet |
Guide 5.1: “Progressive summarization risks condensing numbers, %, dates, customer expectations into vague summaries. Extract transactional facts into a persistent ‘case facts’ block; trim verbose tool outputs to relevant fields; place key summaries at the beginning + explicit section headers.”
- Use a model with a larger context window
- Keep a case-facts block with the promised amount and date, included in every request
- Summarize the chat more often
- Ask the customer to repeat the amount
Three traps that catch people
1 · “Improve the prompt” for a rule that must always hold
If the question gives a failure rate (like 12%) for something involving money or identity, the answer is code: a hook or a prerequisite gate. Stronger wording and more examples are the runner-up.
2 · Sentiment and confidence sound smart
Options that escalate on an angry customer or on Claude's own confidence score look sensible, but the guide calls both unreliable. Look for explicit criteria with examples.
3 · Wrong tool picked? That's the MCP page
When the agent calls get_customer instead of lookup_order, the first fix is the tool descriptions. That's MCP map 3. A failed tool call is MCP map 4.
How to study this
Related: Domain 1 hub · Domain 5 hub · Multi-agent mind maps · Extraction mind maps · Claude Code mind maps · MCP mind maps · all traps · CCA-F start here