Smith: A Workflow-First Multi-Agent Orchestration Framework in the Ruby Ecosystem

Smith is a workflow-first multi-agent orchestration library for Ruby developers, built on RubyLLM. It addresses the transition of AI agent systems from "prompt engineering" to maintainable application systems, offering production-grade features such as explicit state machines, budget control, tool strategies, and persistence hooks. By extracting orchestration logic from prompts into Ruby code, it helps build complex AI systems that are predictable, debuggable, and maintainable.

The name Smith comes from Agent Smith in The Matrix, embodying six system design principles: control, execution, replication, propagation, coordination, and pressure containment. Its origin stems from the increasing complexity of AI agent systems, where pure prompt engineering can no longer meet production environment needs—requiring a shift from "telling the model what to do via prompts" to system-level capabilities like explicit control flows, repeatable debugging, tool guardrails, and budget enforcement.

Core Design Principles: Explicit over implicit (orchestration logic extracted from prompts into Ruby code), workflow-first (workflows as first-class citizens).

Technical Components: Context (manages persistent state, supports pause/resume), Agent (encapsulates RubyLLM, reusable with dynamically generated instructions), Workflow (defines execution flow via state machine pattern, declarative design for ease of analysis and validation). For example, Context uses persist to define persistent data, and Agent uses register_as for registration and reuse.

What Smith Is: A workflow-first agent orchestration library in the Ruby ecosystem, built on RubyLLM, executed in-process, suitable for application-level orchestration with explicit state and control.

What Smith Is Not: Not a managed runtime, not a persistent workflow engine, not a task queue, not a billing-level cost system, and does not replace the application's persistence, retry, or deployment platform. The host application is responsible for implementing persistence, task execution, etc.

Supported System Patterns: Single-agent guardrail systems, sequential multi-step processes, classifier routing, bounded parallel fan-out, nested workflows, generator/evaluator loops, etc.

Practical Value: Compared to ad-hoc prompt branching and ambiguous error handling without Smith, using Smith provides explicit workflow states, clear success/failure routing, inspectable step logs, and extension points like budget and tools.

Status: Under active development, not yet published to RubyGems, can be imported via local path or Git. Provides toolchain like smith doctor (offline/online validation), smith install (generate configuration), and supports Rake tasks and generators for Rails applications.

Ecosystem Value: Fills the gap of multi-agent frameworks in the Ruby ecosystem, allowing Ruby teams to build production-grade AI systems without switching tech stacks, representing the trend of AI applications moving from "demonstration" to "production".