AI Orchestration Framework for Durable Coding Agent Workflows

Smithers is an AI orchestration framework designed for executing long-horizon, multi-step coding agent tasks reliably. It provides durable workflows that survive crashes, enable human approvals, and offer full observability, allowing agents to perform complex, iterative work across various models and environments. It's ideal for scenarios requiring robust, inspectable, and recoverable agent-driven code modifications.

How It Works

Smithers employs a durable execution model where every completed task is persisted to SQLite, ensuring runs can resume from the last checkpoint after any interruption. It supports a flexible "any agent, any model" architecture, allowing seamless integration with diverse coding agents (like Claude Code, Codex) and AI models via SDKs, without requiring workflow rewrites. The framework enhances output quality through structured review loops, human approval gates, and automated evaluations, facilitating real-world development demands.

Quick Start & Requirements

• Installation: Integrate with coding agents using bunx smithers-orchestrator skills add and bunx smithers-orchestrator mcp add. For CLI-driven workflows, use bunx smithers-orchestrator init followed by bunx smithers-orchestrator workflow run.
• Prerequisites: Requires compatible coding agents (e.g., Claude Code, Codex, Cursor, Copilot, Pi) and bunx. Supports various sandboxing environments including local, gVisor, Kubernetes, and managed remote providers.
• Documentation: Full details available at smithers.sh. Examples and runnable workflows are in the examples/ directory.

Highlighted Details

• Durable Execution: Workflows persist state to SQLite, enabling crash recovery, retries, and replayability.
• Agent/Model Agnostic: Seamlessly swap AI models and agent harnesses within workflows.
• Enhanced Quality Control: Features human approvals, review loops, and automated evaluations for structured development.
• Observability: Integrates with Prometheus and OpenTelemetry for metrics and tracing.
• Advanced Features: Includes prompt optimization, hot-reloading of workflow components mid-run, and robust sandboxing for isolated execution.
• Security: Built-in controls like human approvals, inspectable/reversible runs, and isolated sandboxes mitigate risks for agents modifying repositories.

Maintenance & Community

The provided README does not detail specific contributors, sponsorships, or community channels (e.g., Discord, Slack).

Licensing & Compatibility

Licensed under the MIT License, generally permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

Smithers is not intended for simple, single-prompt interactions or basic chatbot applications. It is specifically designed for complex, multi-step agentic work involving code modification and repository changes. For general-purpose durable backend jobs, alternative solutions like Temporal or Inngest are suggested.