Audience: applications that need local observation or lightweight reactions to Conductor-agent lifecycle events.
Register callback handlers before starting the runtime. Treat callback payloads as potentially sensitive execution data.
CallbackHandler observes agent lifecycle events and supports hooks for messages,
tool calls, results, and failures. Keep callback work fast and non-blocking; move
durable business effects into a tool or workflow task. Use structured logging and
redact credentials and sensitive user data.
Callbacks receive lifecycle notifications without changing the durable workflow unless they explicitly fail the local process. Do not use them as the only record of an audit, notification, or external write: process restarts can interrupt local observers.
Use streaming for caller-visible events and tools for durable side effects.