High resource usage of AgentServer in Docker limits concurrent rooms on single-core 2GB container

[echo1248]

Feature Type

Nice to have

Feature Description

Environment

• Deployment: Docker container
• Container specs: 1 CPU core, 2 GB RAM
• OS / Base image: [ ghcr.io/astral-sh/uv:python3.13-bookworm-slim ]
• Python Agent / Node.js Agent: Python
• Default start command: uv run src/agent.py start

Description

When running the AgentServer in a single-core, 2GB RAM Docker container with default multiprocessing setup:

• Each new Job spawns a separate process (forkserver/spawn)
• CPU and memory usage quickly saturates
• Only ~4-5 concurrent rooms can be handled before new jobs are rejected due to load or OOM

Steps to Reproduce

1. Start AgentServer in Docker with single core, 2GB RAM
2. Dispatch multiple rooms/jobs simultaneously
3. Observe no servers available errors after 4-5 rooms

Current Behavior

• High memory footprint per Job
• Multiple Python processes are spawned per Job (observed via ps -ef | grep python)
• Default load_threshold in production (0.7) triggers job rejection quickly

Expected Behavior

• Support more concurrent rooms on low-resource environments
• Optionally provide guidance or configuration options to optimize memory/CPU usage for small containers

Questions / Suggestions

1. Can the AgentServer use async-only execution without separate process per Job, for low-resource deployments?
2. Are there config flags to reduce per-Job memory usage or limit multiprocessing overhead?
3. Is it possible to tune process pool / job executor / idle processes to better fit single-core, 2GB environments?

Workarounds / Alternatives

No response

Additional Context

No response

[longcw]

you can specify the server to use threading via

from livekit.agents import JobExecutorType

server = AgentServer(job_executor_type=JobExecutorType.THREAD)

[echo1248]

Question about turn_detector and silero in AgentSession

Hi,

I noticed in the code:

Initialize agent session

session = AgentSession[UserData](
userdata=userdata,
# vad=ctx.proc.userdata["vad"],
# turn_detection=MultilingualModel(), # noqa Multilingual turn detection
)

If the vad and turn_detection lines are commented out, what impact would that have? Are these two components (turn_detector and silero) mandatory for the agent to function?

If we run a single agent on a machine with 1 CPU core and 2GB of RAM, and each room has 1 user and 1 LLM, how many rooms can this agent handle simultaneously?

Thanks!

[longcw]

see them in docs https://docs.livekit.io/agents/logic/turns/vad/, https://docs.livekit.io/agents/logic/turns/turn-detector/

basically the agent can run without VAD, but in that case the interruption is based on the STT output, there might be a delay when user interrupts the agent.