💫 stella-light-agent

Deprecated — use stella-v2

This agent is superseded by stella-v2 and is no longer developed. Existing deployments keep running and their saved configurations stay valid, but do not choose it for new work.

If you came here for lower cost, save a stella-v2 configuration with only the experts you need enabled (the Expert Pool node has a per-expert on/off). You get comparable per-turn cost and keep everything stella-light never grew.

Why it was deprecated

Its pitch was "faster and cheaper". Only half of that survived.

Cheaper still holds — roughly two LLM calls per turn against stella-v2's seven-plus — but that is now a stella-v2 configuration, not a reason for a second codebase.

Faster no longer holds. stella-v2 gained a bridge generator: it starts speaking a short opening beat immediately, while the rest of the pipeline runs behind it. stella-light has no bridge, so it waits in silence for its single call to finish. On perceived latency — the only latency that matters in a spoken conversation — stella-light is now the slower of the two. It also emits no analytics, so that gap is invisible on the dashboard.

Meanwhile the divergence cost was real. The two agents run the same plans, and features kept landing in stella-v2 only: the bridge, analytics, companion mode, persona token resolution. A plan authored with {{persona.name}} read correctly under stella-v2 and printed the raw token under stella-light — and plan prose is spoken aloud.

Migrating

You used stella-light for…	Do this instead
Lower per-turn cost	A stella-v2 configuration with only task_extraction and noise_detection enabled
A simple agent to develop against	stella-v2 with its default configuration
Faster replies	stella-v2 — the bridge makes it start speaking sooner

Nothing about your plans, personas or sessions changes: those are shared, and both agents drive the same state machine through the same SDK.

Overview

stella-light-agent provides a streamlined voice AI pipeline that sacrifices some advanced features for improved performance:

• Lower memory footprint
• Simpler configuration
• Fewer LLM calls per turn

Comparison with stella-agent

Feature	stella-agent	stella-light-agent
STT Quality	High	Good
Response Latency	~2-3s	~1-2s
Memory Usage	512Mi-2Gi	256Mi-1Gi
Tool Calling	Yes	Limited
Progress Tracking	Yes	Basic
Conversation History	Full	Limited

When to Use

Choose stella-light-agent when:

• Development/Testing: Faster iteration cycles
• Simple Conversations: Q&A, basic support
• Resource Constraints: Limited cluster resources
• Cost Optimization: Lower compute costs
• Low Latency Required: Interactive demos

Configuration

Environment Variables

Variable	Description	Default
OPENAI_API_KEY	OpenAI API key	Required
OPENAI_MODEL	Model to use	gpt-4o-mini
STT_PROVIDER	Speech-to-text provider	sherpa
TTS_PROVIDER	Text-to-speech provider	kokoro
MAX_HISTORY	Max conversation turns to keep	5

Pipeline

The light agent uses a simplified pipeline:

Audio In → STT → LLM → TTS → Audio Out

Key differences from stella-agent:

• Minimal preprocessing
• Shorter context window
• Direct response streaming
• Limited tool support

Resource Requirements

Resource	Request	Limit
CPU	100m	500m
Memory	256Mi	1Gi

Data Channel Messages

Similar to stella-agent but with a reduced message set:

// Transcript updates
{
  type: 'transcript_chunk',
  data: {
    text: string,
    is_final: boolean
  }
}

// Agent status
{
  type: 'agent_status',
  data: {
    status: 'listening' | 'speaking'
  }
}

Deployment

Deploy via the API:

curl -X POST http://localhost:3000/sessions/{sessionId}/agents \
  -H "Content-Type: application/json" \
  -d '{
    "role": "conversational-ai",
    "agentType": "stella-light-agent"
  }'

Or via the Frontend UI by selecting "stella-light-agent" from the agent type dropdown.

Performance Tuning

Reduce Latency

1. Use a smaller LLM model (gpt-4o-mini vs gpt-4o)
2. Reduce MAX_HISTORY to minimize context
3. Use local STT/TTS services

Reduce Memory

1. Lower MAX_HISTORY value
2. Disable unused features
3. Use streaming for all responses

Limitations

• Limited Tool Support: Only basic tools available
• Shorter Context: May lose context in long conversations
• Basic Progress Tracking: No detailed todo management
• Simpler Prompts: Less nuanced conversation handling

Upgrading to stella-agent

If you outgrow stella-light-agent:

1. Update the agent type in your deployment
2. Increase resource limits in your pod configuration
3. Add any additional environment variables for new features
4. Update your plans to use advanced features

See Also

• stella-agent - Full-featured agent
• Agents Overview - Agent comparison
• First Agent - Deployment guide