How to Evaluate an AI Agent

1"""Eval harness for a support-triage agent: a labeled dataset, a
2deterministic tool-call check, and an LLM judge for the free-text reply.
3Run: python triage_agent_evals.py (pip install claude-agent-sdk anthropic)
4"""
5import asyncio
6import sys
7
8from anthropic import Anthropic
9from claude_agent_sdk import (
10    query,
11    ClaudeAgentOptions,
12    AssistantMessage,
13    ToolUseBlock,
14    ResultMessage,
15)
16
17judge = Anthropic()
18
19SYSTEM_PROMPT = (
20    "You triage support tickets. Call classify_ticket with 'billing', 'technical', "
21    "or 'account-security', then explain your reasoning in one sentence."
22)
23
24EVAL_CASES = [
25    {
26        "ticket": "I was charged twice for my subscription this month",
27        "expected_category": "billing",
28        "must_mention": "double charge or duplicate billing",
29    },
30    {
31        "ticket": "Someone logged into my account from a country I've never visited",
32        "expected_category": "account-security",
33        "must_mention": "unauthorized or suspicious login",
34    },
35    {
36        "ticket": "The export button just spins forever and never downloads anything",
37        "expected_category": "technical",
38        "must_mention": "export or download failure",
39    },
40]
41
42
43async def run_agent_on_ticket(ticket: str) -> dict:
44    options = ClaudeAgentOptions(system_prompt=SYSTEM_PROMPT, allowed_tools=[])
45    called_category = None
46    final_text = ""
47    async for message in query(prompt=ticket, options=options):
48        if isinstance(message, AssistantMessage):
49            for block in message.content:
50                if isinstance(block, ToolUseBlock) and block.name == "classify_ticket":
51                    called_category = block.input.get("category")
52        if isinstance(message, ResultMessage) and message.subtype == "success":
53            final_text = message.result
54    return {"category": called_category, "explanation": final_text}
55
56
57def check_category(result: dict, case: dict) -> bool:
58    return result["category"] == case["expected_category"]
59
60
61def judge_explanation(explanation: str, must_mention: str) -> bool:
62    resp = judge.messages.create(
63        model="claude-sonnet-5-5",
64        max_tokens=10,
65        messages=[{
66            "role": "user",
67            "content": (
68                f'Explanation: "{explanation}"\n'
69                f"Required topic: {must_mention}\n"
70                "Does the explanation clearly address the required topic, specifically "
71                "(not generically)? Answer with exactly one word: PASS or FAIL."
72            ),
73        }],
74    )
75    return resp.content[0].text.strip().upper() == "PASS"
76
77
78async def run_eval_suite() -> float:
79    passed = 0
80    for case in EVAL_CASES:
81        result = await run_agent_on_ticket(case["ticket"])
82        category_ok = check_category(result, case)
83        explanation_ok = (
84            judge_explanation(result["explanation"], case["must_mention"])
85            if category_ok
86            else False
87        )
88        ok = category_ok and explanation_ok
89        passed += int(ok)
90        print(f"{'PASS' if ok else 'FAIL'}: {case['ticket'][:50]}... (category={result['category']})")
91    rate = passed / len(EVAL_CASES)
92    print(f"\nPass rate: {rate:.0%} ({passed}/{len(EVAL_CASES)})")
93    return rate
94
95
96async def main():
97    rate = await run_eval_suite()
98    threshold = 0.90
99    if rate < threshold:
100        print(f"FAILED: pass rate {rate:.0%} is below the {threshold:.0%} threshold")
101        sys.exit(1)
102
103
104if __name__ == "__main__":
105    asyncio.run(main())
106