Agente runtime

Execute um agente autônomo a partir do código — sem UI. O runtime executa um loop ReAct: chama ferramentas, observa resultados, decide a próxima ação.

Código

import { createRuntime } from '@agentskit/runtime'
import { researcher } from '@agentskit/skills'
import type { ToolDefinition, Observer, AgentEvent } from '@agentskit/core'

const webSearch: ToolDefinition = {
  name: 'web_search',
  description: 'Search the web for information',
  schema: {
    type: 'object',
    properties: { q: { type: 'string', description: 'Search query' } },
    required: ['q'],
  },
  execute: async (args) => {
    // Replace with real search API
    return `Results for "${args.q}": [1] Paper A, [2] Paper B`
  },
}

const logger: Observer = {
  name: 'console',
  on(event: AgentEvent) {
    if (event.type === 'agent:step') console.error(`[step ${event.step}]`)
    if (event.type === 'tool:start') console.error(`[tool] ${event.name}`)
  },
}

const runtime = createRuntime({
  adapter: yourAdapter,  // openai, anthropic, etc.
  tools: [webSearch],
  observers: [logger],
})

const result = await runtime.run('Research AI safety developments', {
  skill: researcher,
})

console.log(result.content)
// Completed in 2 steps, 1 tool call

Experimente

cd apps/example-runtime
pnpm dev

Defina OPENAI_API_KEY para um provedor real, ou execute sem para modo demo.