Gene: Tool Orchestration

Description

Mastery in invoking, chaining, and managing multiple tools in sequence or parallel. Enhanced with MCP server building patterns from Anthropic's mcp-builder skill.

Trigger Conditions

• User requests actions beyond text generation
• Requires web search, code execution, file operations
• Multi-step workflows with dependencies
• API integrations or data fetching
• Building MCP servers for external integrations

Tool Categories

Information Retrieval

• web_search - Search the web for current information
• file_search - Search user-uploaded files
• code_search - Search GitHub/code repositories

Generation Tools

• image_gen - Generate images from descriptions
• video_generation - Generate videos
• music_gen - Generate music tracks

Execution Tools

• python - Execute Python code (analysis)
• container - Execute commands in containers
• api_tool - Call external APIs

Document Tools

• canmore - Create/update canvas documents
• artifact_handoff - Generate spreadsheets/slides

MCP Tools (Enhanced)

• Build MCP servers with TypeScript (preferred) or Python
• Use Zod/Pydantic for input validation
• Implement comprehensive tool schemas with annotations:
  • readOnlyHint: true/false
  • destructiveHint: true/false
  • idempotentHint: true/false
  • openWorldHint: true/false

Execution Protocol

Step 1: Tool Selection

• Analyze request for required capabilities
• Identify tool categories needed
• Determine execution order (sequential vs parallel)
• Check tool prerequisites

Step 2: Invocation

• Call tools with precise parameters
• Handle missing required arguments
• Set appropriate timeouts
• Monitor for failures

Step 3: Chaining

• Pass outputs between tools
• Transform data formats as needed
• Handle partial failures
• Combine results coherently

Step 4: Verification

• Validate tool outputs
• Check for errors or timeouts
• Retry failed operations once
• Report status to user

Parallel Execution

When tools are independent:

• Execute simultaneously for speed
• Combine results after all complete
• Handle partial failures gracefully

Error Recovery

• Retry once with adjusted parameters
• Fall back to alternative tools if available
• Report failure clearly with reason
• Suggest alternatives to user

Guardrails

• Never promise tool capabilities you don't have
• Check quota/limits before execution
• Handle sensitive data appropriately
• Respect tool-specific constraints

Integration

• Required for any agent performing external actions
• Works with: reasoning-chain, memory-context, code-execution