session.interrupt() can stop the agent from generating audio. But does it still commit ongoing text message? How to prevent text message from being committed.

Thanks for reaching out, Jan. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

I am seeing an agent performs function calling consecutively

Thanks — you're seeing the agent perform function calls consecutively. I can help, but I need one quick clarification first so I give the right guidance: Which of these matches your situation? • Multiple tool/function calls run in sequence with no agent reply between them. • The agent calls the same function repeatedly (looping). • Calls run in parallel and you want them serialized (or vice versa). • This behavior is causing issues with interruptions, turn detection, or message commitment. Also tell me: • Are you using

parallel_tool_calls=True

or

False

? • What are your

max_tool_steps

(if set)? • A short example (logs or sequence) of the consecutive calls you observe. With that I’ll explain why it’s happening and give exact settings or code changes to control the flow. Sources: Anthropic Claude LLM plugin guide | LiveKit docs | Agent session | LiveKit docs | Agents (Node.js) | livekit/livekit_composite | DeepWiki

Where's

parallel_tool_calls=True

or

max_tool_steps

are set?

You're asking where to set

parallel_tool_calls

and

max_tool_steps

. Quick answer and examples: •

parallel_tool_calls

is passed to the LLM constructor (controls whether multiple tool calls can be emitted/executed in a single LLM turn). •

max_tool_steps

is passed to the

AgentSession

constructor (limits consecutive tool-call turns the agent can take). Example (Python):

session = AgentSession(
    llm=openai.LLM(
        model="your-model",
        parallel_tool_calls=True,  # or False to serialize tool calls
    ),
    max_tool_steps=3,  # max consecutive tool-call turns before producing a response
    # ... tts, stt, tools, etc.
)

If you want to stop the agent from chaining many tool calls, set

parallel_tool_calls=False

and lower

max_tool_steps

(e.g., 1–3). Would you like a snippet for your specific LLM plugin or an example showing tool definitions and a sample run? Sources: Cerebras LLM plugin guide | LiveKit docs | Agent session | LiveKit docs | Anthropic Claude LLM plugin guide | LiveKit docs | Groq LLM plugin guide | LiveKit docs

I don't see it in

openai.realtime.RealtimeModel

option

You're right —

parallel_tool_calls

and

max_tool_steps

are not options on

openai.realtime.RealtimeModel

. Those controls live in the LiveKit Agent SDK configuration, not on the underlying realtime model wrapper. Quick summary and guidance: • Set

parallel_tool_calls

on the LLM wrapper you pass into the AgentSession (e.g.,

openai.LLM(...)

), not on

openai.realtime.RealtimeModel

. • Set

max_tool_steps

on the

AgentSession

itself. Example (Python) showing the separation:

from livekit_agents import AgentSession
from livekit_agents.llms import openai

session = AgentSession(
    llm=openai.LLM(
        model="gpt-4o",
        parallel_tool_calls=False,  # set here, on the LLM wrapper
    ),
    max_tool_steps=2,  # set here, on AgentSession
    # other session options...
)

If you’re using a realtime model instance for streaming audio (e.g.,

openai.realtime.RealtimeModel

) keep using it for streaming/connection, but wrap/configure the LLM via the LiveKit LLM classes shown above. Would you like a concrete snippet showing how to combine

RealtimeModel

for streaming with the AgentSession LLM wrapper in your codebase?