Agent Skill Composer

Dynamic tool chaining system for AI agents to build complex workflows from atomic skills.

Overview

Agent Skill Composer is a powerful framework designed to enable AI agents to compose multiple specialized "skills" into complex, multi-step workflows. It handles input/output validation, data mapping between steps, conditional execution, retries, and state persistence.

Key Features

• Atomic Skills: Define skills with strongly-typed Zod schemas for inputs and outputs.
• Dynamic Chaining: Map outputs from one step to inputs of another using simple JSON paths.
• Validation: Automatic runtime validation of all inputs and outputs using Zod.
• Conditional Logic: Branching workflows based on variable states.
• Resilience: Configurable retry logic for individual steps.
• Persistence: Built-in state management to track and recover workflow progress.
• CLI Interface: Easy-to-use command-line interface for listing skills and running workflows.

Installation

bun install

Usage

List Available Skills

bun src/index.ts list

Run Example Workflow

bun src/index.ts run-example --text "AI agents are changing the world of automation."

Architecture

The system is composed of several core modules:

1. SkillRegistry: Stores and manages SkillDefinition objects.
2. ExecutionEngine: The heart of the system that orchestrates workflow execution.
3. StatePersister: Manages the persistence of workflow state to disk.
4. WorkflowDefinition: A structured representation of a multi-step process.

Example Skill Definition

registry.register({
  name: "text-transformer",
  description: "Transforms text to uppercase or lowercase",
  inputSchema: z.object({
    text: z.string(),
    mode: z.enum(["uppercase", "lowercase"])
  }),
  outputSchema: z.object({
    result: z.string()
  }),
  execute: async (input) => {
    return {
      result: input.mode === "uppercase" ? input.text.toUpperCase() : input.text.toLowerCase()
    };
  }
});

License

MIT