Callbacks

Subclass CallbackHandler and override the lifecycle hooks you care about. Each hook runs as a server-registered worker.

import { CallbackHandler } from '@io-orkes/conductor-javascript/agents';

class Logger extends CallbackHandler {
  async onAgentStart(agentName: string, prompt: string) { console.log('[start]', agentName, prompt); }
  async onToolStart(agentName: string, toolName: string, args: unknown) { console.log('[tool]', toolName, args); }
  async onAgentEnd(agentName: string, result: unknown) { console.log('[end]', agentName); }
}

const agent = new Agent({ name: 'a', model, instructions: '…', callbacks: [new Logger()] });

Hooks: onAgentStart, onAgentEnd, onModelStart, onModelEnd, onToolStart, onToolEnd.

Expected behavior and failures

Keep callback work fast and non-blocking — move durable business effects (writes, notifications, audit records) into a tool() or a workflow task instead. Callbacks observe lifecycle events without changing the durable workflow unless they explicitly throw; don't rely on them as the sole record of an audit or external write, since a process restart can interrupt a local observer mid-callback. Treat callback payloads (messages, tool args/results) as potentially sensitive execution data — redact credentials before logging.

Next steps

Use streaming for caller-visible progress events and tools for durable side effects.