Bulkhead τ Meta-Domain

Operational Modes

Meta-Tool vs Direct (500+)

Two Operational Modes

ShowcaseAgent supports two modes optimized for different LLM capabilities.

Recommended for Local LLMs

Meta-Tool Mode

Meta-Tools

One meta-tool per domain. LLM makes one decision: "which domain?"

  • Reduced context window pressure
  • Simpler tool selection
  • Natural language queries per domain
  • Better for Ollama, small models
For Cloud LLMs

Direct Mode

500+ Tools

All individual tools flattened. LLM sees every tool and selects directly.

  • Full tool granularity
  • Precise tool selection
  • Requires large context window
  • Better for Claude, GPT-4

Meta-Tool Mode Example

# Meta-Tools Available to LLM query_tennisagent_data(query: str) "Natural language query for tennis sensor analysis" query_portfolio_data(query: str) "Natural language query for investment portfolio" query_optiver_data(query: str) "Natural language query for trading ML" query_wq_data(query: str) "Natural language query for WQ data science" query_stan_data(query: str) "Natural language query for Stats 315A" query_americaneconomy_data(query: str) "Natural language query for U.S. macro and fiscal analysis" query_globaltemperature_data(query: str) "Natural language query for climate and emissions analysis" ... 7 more meta-tools # User Query Flow User: "show my forehand consistency analysis" | v LLM selects: query_tennisagent_data("show my forehand consistency analysis") | v TennisAgent processes query internally | v Result returned to user

Direct Mode Example

# 500+ Individual Tools Available to LLM # TennisAgent tools analyze_match_consistency(activity_id: str, shot_type: str) detect_fatigue_patterns(activity_id: str) visualize_zepp_impact(start_date: str, end_date: str) train_stroke_classifier(model_type: str) ... # Portfolio tools (70) list_holdings() get_tax_summary() find_harvest_opportunities() calculate_options_income() ... # ... 394 more tools from other domains # User Query Flow User: "show my forehand consistency analysis" | v LLM selects: analyze_match_consistency(activity_id="latest", shot_type="forehand") | v Tool executes directly | v Result returned to user

Mode Comparison

Aspect Meta-Tool Mode Direct Mode
Tool Count one per domain 500+ tools
Context Usage Low (~2K tokens) High (~50K tokens)
LLM Decision "Which domain?" "Which specific tool?"
Query Interface Natural language Structured parameters
Best For Local LLMs (Ollama, llama.cpp) Cloud LLMs (Claude, GPT-4)
Routing ShowcaseAgent routes LLM routes directly
Granularity Domain-level Tool-level

Experimental Variations

ShowcaseAgent includes 3 variations for comparative studies.

6

V1: Browse

Read-only catalog exploration

V2

Meta-Tool

Aggregated domain interfaces

500+

V3: Direct

All individual tools

Variation 1: Browse Mode (6 tools)

Tool Description
list_domains List all available domains
get_domain_info Get detailed info for a specific domain
list_domain_tools List tools available in a domain
search_tools Search for tools by keyword
get_tool_details Get detailed tool documentation
get_domain_stats Get aggregate statistics

Variation 2: Meta-Tool Mode

Tool Domain
query_tennisagent_data TennisAgent
query_optiver_data Optiver
query_portfolio_data Portfolio
query_wq_data WQ
query_qci_data QCiAgent
query_parable_data ParableAgent
query_stan_data Stan
query_yieldmodel_data YieldModel
query_aiwq_data AI_WQ
query_ppr_data PPR_Agent
query_garmin_data GarminAgent
query_golf_data GolfAgent
query_americaneconomy_data AmericanEconomy
query_globaltemperature_data GlobalTemperature