Thanks — I’m responding to your question about how Agent deployments work for development vs production releases for your app. High-level summary • The same agent code can be used for both development and production; differences come from the runtime mode, environment variables, and deployment method. • Development focuses on fast iteration and local testing; production focuses on reliability, scaling, and observability. Development workflows (fast iteration) • Run locally in console mode to test logic and get immediate feedback (Python agents can run in

console

mode). • Use

dev

mode to connect your agent to LiveKit Cloud so it’s reachable from anywhere — useful for testing with real clients or the Agents Playground. • Alternatively run a local LiveKit server for full local control; the agent code doesn’t need changes when you later deploy to production. • Keep secrets and keys (OpenAI, Deepgram, LiveKit credentials) in local env vars or a local secrets file while developing. Production workflows (scale, reliability, observability) • Use the production runtime mode (for Python agents the typical command is `uv run agent.py start`; Node agents need build + start, e.g.,

pnpm build

then

pnpm start

). • Use

lk agent create

to generate deployment artifacts (for example

Dockerfile

and

livekit.toml

) and to deploy to LiveKit Cloud. • LiveKit Cloud runs agents in containers across global data centers, handling provisioning, load balancing, metrics, and rollbacks. • Production uses environment variables injected at deploy time for required secrets (

OPENAI_API_KEY

,

LIVEKIT_API_KEY

,

LIVEKIT_URL

, etc.). Use separate LiveKit projects for staging and production to isolate traffic. Scaling and updates • Agent instances scale elastically: when capacity approaches limits, new instances are started to serve sessions. • Updates are zero-downtime: existing sessions are drained gracefully while new sessions are routed to updated instances. • Rollbacks can be performed quickly if you need to revert a release. CI/CD and release automation • Integrate with GitHub Actions (LiveKit provides a deploy action) to automate deployments on branch pushes. • Use separate branches/environments (development, staging, production) and repository secrets to manage different LiveKit projects and env vars. • Typical flow: push to

main

or

production

branch -> CI builds the agent, injects secrets for the target environment -> deploys to LiveKit Cloud. Configuration and metadata • Manage agent metadata and routing using

livekit.toml

in your project root (subdomain, agent id, etc.). • Keep environment-specific configuration out of source control (use CI secrets, environment variables, or your secrets manager). Observability • Use agent insights and the LiveKit dashboard to monitor session analytics, latency, uptime, resource usage, and metrics like TTFT/TTFB and turn latency. • Ensure logging and metrics are enabled for production agents so you can debug and measure performance. If you’d like, I can help with one of the following next steps: • Walk you through a specific local dev run (console vs dev) for Python or Node agents. • Create a checklist for promoting an agent from staging to production (build, env vars, deploy, smoke tests). • Help set up a GitHub Actions workflow example for automated deploys. Which of these would be most useful right now? Sources: Voice AI quickstart | LiveKit Documentation | An open source stack for real-time multimodal AI | Building the all-in-one platform for voice AI agents | Deploy and scale agents on LiveKit Cloud | Self-hosted deployments | LiveKit Documentation