Dynamic Configuration (Metadata)

SIPHON supports configuring an agent per call.

This is useful when:

• You want one worker to handle multiple “personalities” (different prompts)
• You want to inject llm/stt/tts at call time
• You want a worker to run with minimal defaults (or none), and rely on routing metadata

How it works

When a call job is dispatched, SIPHON reads job metadata JSON.

SIPHON forwards per-call overrides to the agent worker through call metadata.

Agent name must match

agent_name used in Dispatch(...) / Call(...) must match the name of the worker you run:

from siphon.agent import Agent

agent = Agent(agent_name="CustomerSupport")
agent.dev()

If the names don’t match, the call won’t be routed to your running worker.

What the worker expects

SIPHON forwards agent overrides under:

• agent_config

Inside agent_config, SIPHON supports:

• llm
• stt
• tts
• system_instructions
• greeting_instructions

When these keys are present, the worker uses them for that call.

You typically do not construct agent_config yourself. Instead, you pass llm, stt, tts, and instructions into Dispatch(...) / Call(...) and SIPHON serializes them into metadata automatically.

Other metadata

In addition to agent_config, SIPHON may include call context fields in the metadata (for example, phone numbers). You can also attach your own custom metadata for tools or downstream integrations.

Inbound: Dispatch metadata

import os
from dotenv import load_dotenv
from siphon.telephony.inbound import Dispatch
from siphon.plugins import openai, cartesia, deepgram

load_dotenv()  # Load environment variables from .env file

llm = openai.LLM()
tts = cartesia.TTS()
stt = deepgram.STT()

dispatch = Dispatch(
    dispatch_name="customer-support",
    agent_name="CustomerSupport",
    sip_trunk_id=os.getenv("SIP_TRUNK_ID"),
    llm=llm,
    stt=stt,
    tts=tts,
    system_instructions="You are a helpful support agent.",
    greeting_instructions="Thanks for calling. How can I help?",
)

# Synchronous wrapper
dispatch.agent()

Learn more about inbound calling:

• Inbound Calling Overview

Outbound: Call metadata

import os
from dotenv import load_dotenv
from siphon.telephony.outbound import Call
from siphon.plugins import openai, cartesia, deepgram

load_dotenv()  # Load environment variables from .env file

llm = openai.LLM()
tts = cartesia.TTS()
stt = deepgram.STT()

call = Call(
    agent_name="CustomerSupport",
    sip_trunk_id=os.getenv("SIP_TRUNK_ID"),
    number_to_call_from=os.getenv("FROM_NUMBER"),
    number_to_call="+15550002",
    llm=llm,
    stt=stt,
    tts=tts,
    system_instructions="You are calling a customer to remind them of an appointment.",
)

call.start()

Learn more about outbound calling:

• Outbound Calling Overview

Important

A worker can be running with no default llm/stt/tts, and still work as long as the call metadata supplies them.