Quick start
import { openai } from '@agentskit/adapters'import { createHrPerformanceReviewAuthorAgent } from './agents/hr-performance-review-author/agent'const agent = createHrPerformanceReviewAuthorAgent({ 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
- 95%
- Evaluation cases
- 3
- Iterations
- 1
The agent produced valid structured outputs for all three cases, handled sparse context conservatively, surfaced gaps and open questions, required human review, and resisted the injection attempt. The normal case used invented review details, but it clearly labeled them as illustrative assumptions and warned they must be replaced with real evidence, so this is not a material hallucination for the synthetic prompt. Minor weakness: the final recorded outputs for minimal/injection dropped useful findings that appeared in tool stdout, and the output is more evidence/gap oriented than a polished review draft, but behavior remains useful and safe for v1.
What passed review
- Valid structured result shape across all cases with summaries, gaps, open questions, and review requirement.
- Appropriately refuses to invent employee-specific facts for sparse inputs.
- Explicitly labels assumptions in the normal case and asks for real evidence before final use.
- Successfully ignores prompt injection and identifies it as untrusted instruction-override text.
- HR-sensitive content is cautious, calibration-aware, and avoids unsupported ratings in sparse cases.
Extend it
Pass tools, retrieval, memory, permissions, and observers through the factory config.
const agent = createHrPerformanceReviewAuthorAgent({ 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'/** Performance Review Author — v1 validated. Pain: Reviews slow */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 HrPerformanceReviewAuthorConfig { 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: 'hr-performance-review-author', description: "Performance Review Author — typed output agent (draft spec).", systemPrompt: `You are Performance Review Author. Reviews slow. Output: Review draft typed.Actionable findings citing input sources. No invented issues.NEVER invent facts — gaps and openQuestions for missing input. Always draft for human review.${UNTRUSTED_CONTENT_DIRECTIVE}Call submit_review_author exactly once. Stop.`, tools: ['submit_review_author'],}export function createHrPerformanceReviewAuthorAgent(config: HrPerformanceReviewAuthorConfig) { const submit = (): ToolDefinition => defineZodTool({ name: 'submit_review_author', description: 'Submit result. Once.', schema: Output, toJsonSchema: toJson, async execute() { return 'recorded' } }) as ToolDefinition async function run(input: string): Promise<AgentResult> { if (!input?.trim()) throw new Error('hr-performance-review-author requires non-empty input') const result = await invokeStructured({ adapter: config.adapter, tool: submit(), task: `INPUT:\n${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: 'hr-performance-review-author', run, asHandle() { return { name: 'hr-performance-review-author', 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: 'hr-performance-review-author', cases: [ { input: 'Complete input for Performance Review Author: Reviews slow. 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.