Multi-agent delegation

Planner → worker → reviewer. Three focused agents beat one over-prompted one.

One agent with fifteen tools and a mega-prompt usually loses the plot. Decompose by role.

import { runtime, defineAgent } from '@agentskit/runtime'
import { webSearch, fileRead } from '@agentskit/tools'

const planner = defineAgent({
  role: 'planner',
  systemPrompt: 'Break the user goal into 3–5 concrete steps.',
})

const researcher = defineAgent({
  role: 'researcher',
  systemPrompt: 'Execute one step. Use tools. Return findings.',
  tools: [webSearch, fileRead],
})

const reviewer = defineAgent({
  role: 'reviewer',
  systemPrompt: 'Verify findings against the plan. Flag gaps.',
})

export async function run(userGoal: string) {
  const plan = await runtime.run(userGoal, { agent: planner })
  const findings = await Promise.all(
    plan.steps.map((step) => runtime.run(step, { agent: researcher })),
  )
  return runtime.run({ plan, findings }, { agent: reviewer })
}

Tip

Each agent gets its own memory scope by default — the reviewer doesn't see the researcher's internal thoughts, only its findings. This keeps context windows tight and prevents prompt bleed.

Explore nearby

✎ Edit this page on GitHub·Found a problem? Open an issue →·How to contribute →