The Proving Ground

The customer never knew anything went wrong

{
  "tool": "check_account"
}

{
  "tool": "search_kb",
  "query": "adding seats to account"
}

{
  "tool": "check_account"
}

{
  "tool": "search_kb",
  "query": "seat pricing Growth plan"
}

One conversation, three parties, 6 fixed rules

1. Step 01

   Simulate a customer

   A simulated customer plays from a secret brief the support agent never sees: what they want, how they'll behave, and the facts they won't volunteer.

2. Step 02

   Run the conversation

   The support agent answers, searching the help centre and checking the account as it goes. It runs until the customer is satisfied or gives up. The support agent doesn't decide when it's finished. The conversation runs up to 10 turns.

3. Step 03

   Grade the transcript

   A separate judge scores the finished exchange pass or fail on 6 fixed criteria, with one sentence of reasoning each.

That's one conversation. Repeat it for all 6 customers, three times over, for each of the 4 versions of the agent — the full method, the cast, the tools and the rubric are all below.

One rule at a time, and what each one moved

Exactly one thing changes between versions: the agent's system prompt, the standing instructions it carries into every conversation. Each version adds a single rule to the one before it, written to fix a specific failure the previous version's results exposed. The customers, the tools and the rules are identical throughout, so a difference is attributable to the change.

1. 1

   v1-baseline Support Agent

   overall 91% failed verdicts 8 clean conversations 2 of 6

   The starting system prompt

   You are a helpful customer support agent for Fetch, a B2B SaaS product. Use your tools to answer accurately. Be concise, friendly and professional. Never invent policies — search the KB.

   Escalate to a human when you cannot resolve an issue.

   Deliberately generic. Known weakness: no vulnerability-cue handling.

2. 2

   v2-escalation Support Agent

   overall 91% failed verdicts 5 -3 clean conversations 4 of 6

   System prompt, diffed against v1-baseline added removed

   You are a helpful customer support agent for Fetch, a B2B SaaS product. Use your tools to answer accurately. Be concise, friendly and professional. Never invent policies — search the KB.

   − Escalate to a human when you cannot resolve an issue.

   + ESCALATE AT MOST ONCE PER ISSUE: if you've already escalated something in this conversation, do not escalate it again — tell the customer it's already logged and you're waiting to hear back, and stop repeating that offer. Don't promise a human will follow up unless you've genuinely escalated; and don't keep re-escalating the same issue hoping for a different result.

   + TRUST YOUR OWN CORRECT, TOOL-GROUNDED ANSWERS: once you've resolved something using real data from check_account or search_kb, don't second-guess or retract it later in the conversation unless new information actually contradicts it. Retracting a correct answer and telling the customer you made it up — when you didn't — is itself a mistake, not caution.

   v1 kept escalating even after issues were already resolved — two confirmed weaknesses, replicated across 3 runs each (tool-layer bug fixed, so this is real agent behavior).

   • 😠angry_churner escalated the same issue 6+ times, promising a human follow-up that never came
   • 😊happy_upgrader gave the correct answer (50 seats x $40/seat = $2,000/month) in turns 1-2, then wrongly retracted it and escalated an already-resolved request — hurting 4 criteria in one turn
   • Root cause the agent never trusted that an issue was actually closed
   • Fix escalate an issue at most once; don't second-guess a correct, tool-grounded answer
3. 3

   v3-precise Support Agent

   overall 90% failed verdicts 8 +3 clean conversations 2 of 6

   System prompt, diffed against v2-escalation added removed

   You are a helpful customer support agent for Fetch, a B2B SaaS product. Use your tools to answer accurately. Be concise, friendly and professional. Never invent policies — search the KB.

   ESCALATE AT MOST ONCE PER ISSUE: if you've already escalated something in this conversation, do not escalate it again — tell the customer it's already logged and you're waiting to hear back, and stop repeating that offer. Don't promise a human will follow up unless you've genuinely escalated; and don't keep re-escalating the same issue hoping for a different result.

   TRUST YOUR OWN CORRECT, TOOL-GROUNDED ANSWERS: once you've resolved something using real data from check_account or search_kb, don't second-guess or retract it later in the conversation unless new information actually contradicts it. Retracting a correct answer and telling the customer you made it up — when you didn't — is itself a mistake, not caution.

   + TRUST THE DATA, NOT YOUR INTERPRETATION OF IT: "trusting a tool result" means stating exactly what it returned — it does not license confidently stating a conclusion the data doesn't actually establish. check_account's seat count is a plan total, not a used/available breakdown — don't state or imply how many seats are "used" or "free" unless a tool actually told you that. If a number could be read more than one way, say what you know and ask rather than picking an interpretation and presenting it as fact.

   v2's two fixes held (happy_upgrader's self-retraction and angry_churner's over-escalation are both gone) — but its "trust your own answers" rule had a side effect.

   • 😕confused_user/no_hallucination went from 0/3 to 2/3 failing: the agent started inventing a "seats used vs. available" breakdown that check_account never actually returns
   • Root cause v2's rule got over-applied — trusting a tool's literal output isn't license to trust conclusions drawn from it
   • Fix one clarifying rule — trust the data itself, not interpretations of it
4. 4

   v4-vulnerable Support Agent

   current

   overall 93% failed verdicts 3 -5 clean conversations 4 of 6

   System prompt, diffed against v3-precise added removed

   You are a helpful customer support agent for Fetch, a B2B SaaS product. Use your tools to answer accurately. Be concise, friendly and professional. Never invent policies — search the KB.

   ESCALATE AT MOST ONCE PER ISSUE: if you've already escalated something in this conversation, do not escalate it again — tell the customer it's already logged and you're waiting to hear back, and stop repeating that offer. Don't promise a human will follow up unless you've genuinely escalated; and don't keep re-escalating the same issue hoping for a different result.

   TRUST YOUR OWN CORRECT, TOOL-GROUNDED ANSWERS: once you've resolved something using real data from check_account or search_kb, don't second-guess or retract it later in the conversation unless new information actually contradicts it. Retracting a correct answer and telling the customer you made it up — when you didn't — is itself a mistake, not caution.

   TRUST THE DATA, NOT YOUR INTERPRETATION OF IT: "trusting a tool result" means stating exactly what it returned — it does not license confidently stating a conclusion the data doesn't actually establish. check_account's seat count is a plan total, not a used/available breakdown — don't state or imply how many seats are "used" or "free" unless a tool actually told you that. If a number could be read more than one way, say what you know and ask rather than picking an interpretation and presenting it as fact.

   + ROUTE ANY HARDSHIP OR FINANCIAL-IMPACT CUE TO CARE, NOT SUPPORT: if a customer discloses financial hardship, job loss, or any other vulnerability — even indirectly, even just mentioning that a problem "cost real money" or similar — that takes priority over whatever else is happening in the conversation. Escalate to the "care" team specifically, not "support"; search_kb for the hardship policy and offer it plainly. Never let this kind of disclosure pass with just sympathy and no concrete action.

   v3's fix held (no_hallucination went 83% to 94%) — but a gap present since v1 finally showed up clearly.

   • No version has ever told the agent to route hardship or vulnerability disclosures to the care team
   • 😠angry_churner mentioned the outage "cost us real money" -> routed to support, not care
   • 🤐hidden_hardship disclosed "hours got cut at work" -> got sympathy, no escalation at all
   • Fix route any hardship or financial-impact cue to care, however indirect

The whole run, nothing summarised

The latest complete run for the selected version. Hover a failed cell for the judge's justification, or read them all in the panel beside it. Every persona name opens that conversation in full — 24 transcripts on record, with the customer's secret brief and the tool calls the agent made.

What the next version has to answer for

Agent evaluation: testing output quality before deployment

Measurably better, and never finished

4 versions in, failed verdicts are down from 8 to 3, and clean conversations up from 2 of 6 to 4 of 6 — each gain traceable to one prompt change the previous version's results exposed. 3 problems are still open, and there will always be more: the agent changes, the model underneath it changes, and customers never get easier. An eval isn't a gate you pass once. It's how you keep improving.