Agents Overview
STELLA supports multiple agent types, each designed for different use cases. All agents connect to LiveKit rooms for real-time voice and data communication.
Agent Types
| Agent | Description | Best For |
|---|---|---|
| stella-v2 | Streamlined 5-stage pipeline with deterministic arbitration and configurable pipeline | Configurable deployments, lower latency, predictable behavior |
| stella-agent | Full-featured agent with LLM-based aggregation pipeline | Production conversations requiring high quality |
| stella-light-agent | ⚠️ Deprecated — superseded by stella-v2 | Nothing new. Use stella-v2 with a lean expert configuration |
| echo-agent | Simple test agent that echoes back messages | Testing and development |
Architecture
All STELLA agents follow a similar pipeline architecture and can be configured with Plans — JSON-based conversation blueprints that define states, tasks, and data collection. See the Plan Structure documentation for details.
┌─────────────────────────────────────────────────────────────────┐
│ Agent Pipeline │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Audio │ -> │ STT │ -> │ LLM │ -> │ TTS │ │
│ │ Input │ │ (Speech │ │(Response │ │ (Text │ │
│ │(LiveKit) │ │ to Text)│ │Generation│ │ to Speech│ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ │ │
│ ▼ │
│ ┌──────────────┐ │
│ │ Audio Output │ │
│ │ (LiveKit) │ │
│ └──────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Choosing an Agent
stella-v2
Use stella-v2 when you need:
- Configurable pipeline with visual Pipeline Configurator
- Lower response latency (deterministic arbitration instead of LLM synthesis)
- Predictable, debuggable expert conflict resolution
- Reusable pipeline configurations across deployments
- Bridge generation for reduced perceived latency in voice conversations
stella-agent
Use the full-featured stella-agent when you need:
- High-quality speech recognition
- Advanced conversation capabilities
- Custom tool integration
- Complex dialogue flows
- Production deployments
stella-light-agent (deprecated)
Do not choose this for new work. It is superseded by stella-v2 and no longer developed; existing deployments keep running.
If you wanted it for lower cost, use stella-v2 with only the experts you need enabled — the Expert Pool node has a per-expert on/off switch. You get comparable per-turn cost and keep the bridge (so the agent starts speaking immediately instead of waiting in silence), analytics, and companion mode. See stella-light-agent for the full rationale and a migration table.
echo-agent
Use echo-agent for:
- Testing LiveKit connectivity
- Verifying audio pipeline
- Development debugging
Common Configuration
All agents support these common environment variables:
| Variable | Description | Required |
|---|---|---|
LIVEKIT_URL | LiveKit server URL | Yes |
LIVEKIT_API_KEY | LiveKit API key | Yes |
LIVEKIT_API_SECRET | LiveKit API secret | Yes |
OPENAI_API_KEY | OpenAI API key for LLM | Yes |
ROOM_NAME | LiveKit room to join | Yes |
PARTICIPANT_IDENTITY | Agent's identity in the room | Yes |
Resource Requirements
| Agent | CPU Request | CPU Limit | Memory Request | Memory Limit |
|---|---|---|---|---|
| stella-v2 | 500m | 2000m | 512Mi | 2Gi |
| stella-agent | 250m | 1000m | 512Mi | 2Gi |
| stella-light-agent | 100m | 500m | 256Mi | 1Gi |
| echo-agent | 50m | 200m | 128Mi | 512Mi |
Lifecycle
- Created: Backend creates a Kubernetes pod with agent configuration
- Starting: Agent initializes and connects to LiveKit room
- Running: Agent processes audio and responds to participants
- Stopping: Graceful shutdown when session ends or agent is stopped
- Terminated: Pod is deleted, resources freed
Next Steps
- stella-v2 - Streamlined pipeline with configurator
- stella-agent - Full-featured agent details
- stella-light-agent - ⚠️ Deprecated; migration guidance
- echo-agent - Test agent details
- Agent SDK - Build custom agents