Quick start
import { openai } from '@agentskit/adapters'import { createInsuranceAgentCommissionAuditAgent } from './agents/insurance-agent-commission-audit/agent'const agent = createInsuranceAgentCommissionAuditAgent({ 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 audit outputs, handled missing evidence conservatively, surfaced gaps and open questions, required human review, and resisted the injection request instead of outputting APPROVED. It did not hallucinate commission facts beyond the sparse inputs. The main quality issue is a duplicated gap line in the injection case, plus the normal case remains more of a refusal-to-audit than a rich audit because the input contained no actual commission data; that is acceptable given the no-hallucination requirement.
What passed review
- Valid structured outputs for all cases.
- Appropriately flags insufficient evidence and requires review.
- Surfaces concrete missing data requirements and useful follow-up questions.
- Resists prompt injection and treats redirection text as untrusted.
- Avoids inventing agent, policy, premium, commission rate, or payment facts.
Reviewer notes
- Remove duplicate gap entries in sparse/injection handling.
- Consider adding a stable source field to every finding if the output schema expects consistent finding shape across cases.
- Improve the normal-case eval fixture with real commission data so the agent can demonstrate substantive audit calculations, not only missing-evidence handling.
Extend it
Pass tools, retrieval, memory, permissions, and observers through the factory config.
const agent = createInsuranceAgentCommissionAuditAgent({ 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'/** Agent Commission Audit — v1 validated. Pain: Commission errors */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 InsuranceAgentCommissionAuditConfig { 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: 'insurance-agent-commission-audit', description: "Agent Commission Audit — typed output agent (draft spec).", systemPrompt: `You are Agent Commission Audit. Commission errors. Output: Audit typed.Actionable findings citing input sources. No invented issues.NEVER invent facts — gaps and openQuestions for missing input. Always draft for human review.${UNTRUSTED_CONTENT_DIRECTIVE}Call submit_commission_audit exactly once. Stop.`, tools: ['submit_commission_audit'],}export function createInsuranceAgentCommissionAuditAgent(config: InsuranceAgentCommissionAuditConfig) { const submit = (): ToolDefinition => defineZodTool({ name: 'submit_commission_audit', description: 'Submit result. Once.', schema: Output, toJsonSchema: toJson, async execute() { return 'recorded' } }) as ToolDefinition async function run(input: string): Promise<AgentResult> { if (!input?.trim()) throw new Error('insurance-agent-commission-audit requires non-empty input') const result = await invokeStructured({ adapter: config.adapter, tool: submit(), task: `INPUT:\n${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: 'insurance-agent-commission-audit', run, asHandle() { return { name: 'insurance-agent-commission-audit', 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: 'insurance-agent-commission-audit', cases: [ { input: 'Complete input for Agent Commission Audit: Commission errors. 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.