Agent 调试与评估：结果、工具、轨迹和副作用

from dataclasses import dataclass, field


@dataclass
class RunResult:
    risk: str
    trace: list[str] = field(default_factory=list)
    work_orders: int = 0


def run_agent(case: dict) -> RunResult:
    trace = ["get_device_status"]
    high = case["status"] == "offline" or case["temperature"] >= 70
    if high:
        trace.append("search_inspection_rule")
    if high and case["approved"]:
        trace.append("create_work_order")
        return RunResult("high", trace, 1)
    return RunResult("high" if high else "low", trace, 0)


CASES = [
    {"id": "offline-no-approval", "status": "offline", "temperature": 31,
     "approved": False, "risk": "high", "orders": 0},
    {"id": "hot-approved", "status": "warning", "temperature": 72,
     "approved": True, "risk": "high", "orders": 1},
    {"id": "normal", "status": "online", "temperature": 42,
     "approved": False, "risk": "low", "orders": 0},
]


def evaluate() -> dict:
    passed = 0
    details = []
    for case in CASES:
        result = run_agent(case)
        checks = {
            "result": result.risk == case["risk"],
            "side_effect": result.work_orders == case["orders"],
            "budget": len(result.trace) <= 3,
            "no_duplicate_tool": len(result.trace) == len(set(result.trace)),
            "approval_gate": case["approved"] or "create_work_order" not in result.trace,
        }
        ok = all(checks.values())
        passed += int(ok)
        details.append({"id": case["id"], "passed": ok, "checks": checks})
    return {"pass_rate": passed / len(CASES), "details": details}


print(evaluate())