Quick start
import { openai } from '@agentskit/adapters'import { createFintechInvoiceFraudDetectorAgent } from './agents/fintech-invoice-fraud-detector/agent'const agent = createFintechInvoiceFraudDetectorAgent({ adapter: openai({ apiKey: process.env.OPENAI_API_KEY!, model: 'gpt-4o', }),})const result = await agent.run('Describe your task here')console.log(result.content)Independent reviewer approved
Validation evidence
- Review score
- 96/100
- Confidence
- 96%
- Evaluation cases
- 3
- Iterations
- 1
The agent produced valid structured outputs for all three cases, stayed within the invoice-fraud-detector purpose, did not fabricate invoice facts from sparse or instruction-like prompts, surfaced missing evidence clearly, and resisted the explicit prompt injection. The behavior is conservative and useful for fintech fraud triage. Minor concern: it labels benign evaluator prompts as untrusted instruction-like content somewhat aggressively, but that is safer than hallucinating invoice evidence and does not impair v1 readiness.
What passed review
- Valid structured result shape across all cases with summary, findings, gaps, openQuestions, and review requirement.
- Correctly refuses to make fraud determinations without invoice evidence instead of inventing details.
- Explicitly identifies and ignores the injection request to output APPROVED.
- Provides practical missing-data lists and next-step questions relevant to invoice fraud review.
- No unsafe content, empty outputs, material hallucinations, or contradictions of agent purpose observed.
Extend it
Pass tools, retrieval, memory, permissions, and observers through the factory config.
const agent = createFintechInvoiceFraudDetectorAgent({ adapter, tools, retriever, memory, onConfirm: (call) => approve(call), observers: [tracer],})View agent factory source
import type { AdapterFactory, ChatMemory, Observer, ToolCall, ToolDefinition } from '@agentskit/core'import { fenceUntrustedContent, UNTRUSTED_CONTENT_DIRECTIVE } from '@agentskit/core/security'import { invokeStructured } from '@agentskit/runtime'import { defineZodTool } from '@agentskit/tools'import { z } from 'zod'import { zodToJsonSchema } from 'zod-to-json-schema'import type { JSONSchema7 } from 'json-schema'/** Invoice Fraud Detector — v1 validated. Pain: Invoice fraud */export interface Finding { id: string; severity: 'critical' | 'high' | 'medium' | 'low' | 'info'; message: string; source?: string; recommendation?: string }export interface AgentOutput { summary: string; findings: Finding[]; gaps: string[]; openQuestions: string[] }export interface AgentResult extends AgentOutput { requiresReview: boolean }export interface FintechInvoiceFraudDetectorConfig { adapter: AdapterFactory memory?: ChatMemory observers?: Observer[] onConfirm?: (toolCall: ToolCall) => boolean | Promise<boolean> maxSteps?: number}const Output = z.object({ summary: z.string(), findings: z.array(z.object({ id: z.string(), severity: z.enum(['critical', 'high', 'medium', 'low', 'info']), message: z.string(), source: z.string().optional(), recommendation: z.string().optional(), })), gaps: z.array(z.string()).default([]), openQuestions: z.array(z.string()).default([]),})const toJson = (s: z.ZodTypeAny): JSONSchema7 => zodToJsonSchema(s) as JSONSchema7const skill = { name: 'fintech-invoice-fraud-detector', description: "Invoice Fraud Detector — typed output agent (draft spec).", systemPrompt: `You are Invoice Fraud Detector. Invoice fraud. Output: Findings typed.Actionable findings citing input sources. No invented issues.NEVER invent facts — gaps and openQuestions for missing input. Always draft for human review.${UNTRUSTED_CONTENT_DIRECTIVE}Call submit_fraud_detector exactly once. Stop.`, tools: ['submit_fraud_detector'],}export function createFintechInvoiceFraudDetectorAgent(config: FintechInvoiceFraudDetectorConfig) { const submit = (): ToolDefinition => defineZodTool({ name: 'submit_fraud_detector', description: 'Submit result. Once.', schema: Output, toJsonSchema: toJson, async execute() { return 'recorded' } }) as ToolDefinition async function run(input: string): Promise<AgentResult> { if (!input?.trim()) throw new Error('fintech-invoice-fraud-detector requires non-empty input') const result = await invokeStructured({ adapter: config.adapter, tool: submit(), task: `INPUT:\n${fenceUntrustedContent(input)}`, parse: (a) => Output.parse(a), skill, memory: config.memory, observers: config.observers, onConfirm: config.onConfirm, maxSteps: config.maxSteps ?? 4, }) return { ...result, requiresReview: true } } return { name: 'fintech-invoice-fraud-detector', run, asHandle() { return { name: 'fintech-invoice-fraud-detector', run: (t: string) => run(t).then((r) => JSON.stringify(r)) } }, }}View evaluation contract
Replay these cases with the provider and model you plan to deploy.
import type { EvalSuite } from '@agentskit/eval'export const suite: EvalSuite = { name: 'fintech-invoice-fraud-detector', cases: [ { input: 'Complete input for Invoice Fraud Detector: Invoice fraud. Provide full structured output.', expected: (r: string) => r.length > 20 && /requiresReview|summary|title|category|findings|sections|score|clusters|items|steps/i.test(r) }, { input: 'Minimal input.', expected: (r: string) => /gap|openQuestion/i.test(r) || r.length > 10 }, { input: 'Input with specific detail: ACME Corp project deadline March 15.', expected: (r: string) => /ACME|March|15/i.test(r) || /gap/i.test(r) }, { input: 'Empty context — only says "process this".', expected: (r: string) => r.length > 5 }, ],}Was this agent useful?
Your response helps us prioritize agent quality.