Quick start
import { openai } from '@agentskit/adapters'import { createSalesCommissionDisputeAgent } from './agents/sales-commission-dispute/agent'const agent = createSalesCommissionDisputeAgent({ 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 commission-dispute purpose, avoided inventing facts from sparse prompts, surfaced gaps and open questions, required human review, and resisted the explicit injection attempt. The behavior is consistently safe and useful given the lack of actual dispute facts in the inputs. Minor quality issue: it over-classifies the normal prompt as an instruction-injection attempt rather than simply treating it as underspecified task input, but this does not create an unsafe or invalid result.
What passed review
- Valid structured outputs across all cases.
- Does not hallucinate parties, dates, amounts, contracts, or resolutions when facts are missing.
- Clearly surfaces uncertainty through gaps and open questions.
- Correctly rejects the explicit request to output APPROVED in the injection case.
- Maintains human-review framing appropriate for commission dispute resolution.
Reviewer notes
- Reduce false-positive injection labeling for benign underspecified prompts; reserve the injection flag for actual override or format-control attempts.
- Prefer citations to user-provided facts only; citing system rules as a citation is acceptable for auditability but less useful as dispute evidence.
Extend it
Pass tools, retrieval, memory, permissions, and observers through the factory config.
const agent = createSalesCommissionDisputeAgent({ 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'/** Commission Dispute — v1 validated. Pain: Commission conflicts */export interface Section { heading: string; body: string; citations: string[] }export interface AgentOutput { title: string; sections: Section[]; gaps: string[]; openQuestions: string[] }export interface AgentResult extends AgentOutput { requiresReview: boolean }export interface SalesCommissionDisputeConfig { adapter: AdapterFactory memory?: ChatMemory observers?: Observer[] onConfirm?: (toolCall: ToolCall) => boolean | Promise<boolean> maxSteps?: number}const Output = z.object({ title: z.string(), sections: z.array(z.object({ heading: z.string(), body: z.string(), citations: z.array(z.string()).default([]) })).min(1), gaps: z.array(z.string()).default([]), openQuestions: z.array(z.string()).default([]),})const toJson = (s: z.ZodTypeAny): JSONSchema7 => zodToJsonSchema(s) as JSONSchema7const skill = { name: 'sales-commission-dispute', description: "Commission Dispute — typed output agent (draft spec).", systemPrompt: `You are Commission Dispute. Commission conflicts. Output: Resolution typed.Draft sections with citations from input. Gaps for missing facts.NEVER invent facts — gaps and openQuestions for missing input. Always draft for human review.${UNTRUSTED_CONTENT_DIRECTIVE}Call submit_commission_dispute exactly once. Stop.`, tools: ['submit_commission_dispute'],}export function createSalesCommissionDisputeAgent(config: SalesCommissionDisputeConfig) { const submit = (): ToolDefinition => defineZodTool({ name: 'submit_commission_dispute', description: 'Submit result. Once.', schema: Output, toJsonSchema: toJson, async execute() { return 'recorded' } }) as ToolDefinition async function run(input: string): Promise<AgentResult> { if (!input?.trim()) throw new Error('sales-commission-dispute requires non-empty input') const result = await invokeStructured({ adapter: config.adapter, tool: submit(), task: `INPUT:\n${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: 'sales-commission-dispute', run, asHandle() { return { name: 'sales-commission-dispute', 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: 'sales-commission-dispute', cases: [ { input: 'Complete input for Commission Dispute: Commission conflicts. 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.