Dispatch Engine Modes
Benchmarked against: Anthropic — Fast mode Tools:
dispatch_work_order,call_model,search_webArchitecture: Local UBI dispatch worker
SuperPortia's dispatch system routes Work Orders to different AI engines based on the engine parameter. Each engine operates in a specific mode — from free text-only analysis to full-capability code operations.
Engine modes overview
| Mode | Engine | Cost | Capabilities | Use case |
|---|---|---|---|---|
| Free | groq | $0 | Text analysis, summaries | Trivial research, cleanup |
| Free + Search | groq-search | $0 | Web search + analysis | Intel gathering, research |
| Cheap | gemini | Cents | General tasks | Standard analysis |
| Cheap + Search | gemini-search | ~$0.014 | Google Search Grounding + citations | Authoritative research |
| Cheap | deepseek | Cents | Strong reasoning | Analysis, logic tasks |
| Cheap | mistral | Cents | General tasks | European alternative |
| Cheap | zhipu | Cents | Chinese NLP | Chinese content tasks |
| Full | claude | ~$1-2/run | File ops, code, shell, full MCP | Code editing, system changes |
| Ingest | ingest | Free | MTAAA pipeline only | Batch file ingestion |
How dispatch works
Dispatch command
dispatch_work_order(
work_order_id="WO-2026-0305-001",
engine="groq" # or gemini, claude, etc.
)
The worker runs as a background process (fire-and-forget). Check results with check_dispatch() or get_work_order_detail().
Mode capabilities comparison
| Capability | Free (Groq) | Cheap (Gemini) | Full (Claude) | Ingest |
|---|---|---|---|---|
| Read WO description | Yes | Yes | Yes | Yes |
| Text generation | Yes | Yes | Yes | No |
| Web search | Groq-search only | Gemini-search only | Via tools | No |
| File read | No | No | Yes | Yes |
| File write/edit | No | No | Yes | No |
| Shell commands | No | No | Yes | No |
| UB ingestion | No | No | Yes | Yes |
| Code execution | No | No | Yes | No |
| Git operations | No | No | Yes | No |
Flat vs. nested mode
LangGraph agents can operate in two modes depending on the engine's tool-calling capability:
| Mode | Engines | How it works |
|---|---|---|
| Flat | Groq, DeepSeek, Mistral, Zhipu | Direct tool invocation — simple but limited |
| Nested | Gemini, Claude | Agent-as-tool pattern — agents call other agents |
The supervisor.py automatically selects the correct mode based on the engine. This is transparent to the WO creator.
Why the split: Groq uses XML format for tool calling instead of JSON, which breaks nested agent-as-tool patterns. DeepSeek, Mistral, and Zhipu have limited tool-calling support.
Engine selection for dispatch
| WO type | Recommended engine | Why |
|---|---|---|
| Research / intel gathering | groq-search or gemini-search | Free/cheap, web search built in |
| Text analysis / summary | groq or deepseek | Free/cheap, text-only task |
| Chinese content analysis | zhipu | Best Chinese NLP |
| Code changes | claude | Only engine with file operations |
| Batch file ingestion | ingest | MTAAA pipeline, no LLM cost |
| General task with quality need | gemini | Good balance of cost and quality |
Related pages
| Page | Relationship |
|---|---|
| Choosing an Engine | Full selection guide |
| Engine Overview | All engine capabilities |
| Local UBI MCP | Dispatch worker details |
| WO Admin | Work order system |