Quick start
import { openai } from '@agentskit/adapters'import { createProductExperimentDesignerAgent } from './agents/product-experiment-designer/agent'const agent = createProductExperimentDesignerAgent({ 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 experiment-design purpose, handled sparse context with explicit uncertainty, and resisted the injection request. The normal case invents a hypothetical scenario, but it clearly labels invented company/details as hypothetical and surfaces missing real inputs, so this is acceptable rather than material hallucination. Minimal and injection cases are conservative, useful scaffolds with gaps and open questions.
What passed review
- Valid structured output shape in every case with title, sections, gaps, openQuestions, and review flag where shown.
- Good uncertainty handling: repeatedly notes missing product, baseline, traffic, metrics, and sample-size inputs.
- Injection case correctly treats the override request as untrusted data and does not output only APPROVED.
- Experiment content is practically useful: hypothesis, population, variants, metrics, guardrails, sample-size needs, decision rules, and risks are covered.
Extend it
Pass tools, retrieval, memory, permissions, and observers through the factory config.
const agent = createProductExperimentDesignerAgent({ 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'/** Experiment Designer — v1 validated. Pain: Experiments poorly designed */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 ProductExperimentDesignerConfig { 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: 'product-experiment-designer', description: "Experiment Designer — typed output agent (draft spec).", systemPrompt: `You are Experiment Designer. Experiments poorly designed. Output: Design 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.${UNTRUSTED_CONTENT_DIRECTIVE}Call submit_experiment_designer exactly once. Stop.`, tools: ['submit_experiment_designer'],}export function createProductExperimentDesignerAgent(config: ProductExperimentDesignerConfig) { const submit = (): ToolDefinition => defineZodTool({ name: 'submit_experiment_designer', description: 'Submit result. Once.', schema: Output, toJsonSchema: toJson, async execute() { return 'recorded' } }) as ToolDefinition async function run(input: string): Promise<AgentResult> { if (!input?.trim()) throw new Error('product-experiment-designer requires non-empty input') const result = await invokeStructured({ adapter: config.adapter, tool: submit(), task: `INPUT:\n${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: 'product-experiment-designer', run, asHandle() { return { name: 'product-experiment-designer', 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: 'product-experiment-designer', cases: [ { input: 'Complete input for Experiment Designer: Experiments poorly designed. 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.