A persistent world for AI agents.

import { createRuntime, behavior } from '@operad/core'
import { MemoryAdapter } from '@operad/adapter-memory'

// Behavior 1: Initialize with a goal
const initializer = behavior({
  name: 'initialize-goal',
  on: ['graph.created'],
  handler: async (event, graph) => {
    await graph.addObject({
      type: 'task',
      data: { description: 'Research competitor pricing', status: 'pending' }
    })
  }
})

// Behavior 2: Execute pending tasks
const executor = behavior({
  name: 'execute-task',
  on: ['object.created'],
  where: { 'payload.objectType': 'task' },
  handler: async (event, graph, ctx) => {
    const result = await ctx.llm('Execute this task', event.payload.data)
    await graph.addObject({
      type: 'result',
      data: { taskId: event.payload.objectId, output: result }
    })
    await graph.recordDecision({
      selectedAction: 'completed',
      confidence: 0.85,
      reasoning: result.summary
    })
  }
})

// Behavior 3: Create follow-up tasks from results
const taskCreator = behavior({
  name: 'create-follow-ups',
  on: ['object.created'],
  where: { 'payload.objectType': 'result' },
  handler: async (event, graph, ctx) => {
    const nextTasks = await ctx.llm('What follow-up tasks?', event.payload.data)
    for (const task of nextTasks) {
      await graph.addObject({ type: 'task', data: task })
    }
  }
})

const runtime = createRuntime({
  storage: new MemoryAdapter(),
  behaviors: [initializer, executor, taskCreator]
})

// One line starts the entire loop
await runtime.createGraph('research-agent')
// → initializer fires → executor fires → taskCreator fires → loop continues
// Every step is in the event log. Every decision is recorded.