Stoneforge Launches Open-Source Orchestration Platform for Parallel AI Coding Agents
Key Takeaways
- ▸Stoneforge is an open-source orchestration layer that manages multiple AI coding agents working in parallel on the same repository, preventing context degradation and merge conflicts
- ▸The platform uses a specialized multi-agent architecture with Director (planning), Worker (execution), and Steward (review/merge) roles, with each worker running in isolated git worktrees
- ▸Event-sourced architecture uses JSONL as source of truth and SQLite as a rebuildable cache, providing complete audit logs and surviving corruption
Summary
Developer Adam King has released Stoneforge, an open-source orchestration platform designed to manage multiple AI coding agents working simultaneously on the same codebase. The Apache 2.0-licensed tool addresses the "coordination tax" developers face when running multiple AI assistants like Claude Code, which often leads to context switching, merge conflicts, and degraded context windows. Stoneforge introduces a multi-agent architecture with specialized roles: a Director agent that breaks down goals into tasks, Worker agents that execute in isolated git worktrees, and Steward agents that review and merge completed work.
The platform makes several notable architectural choices, including full event-sourcing with JSONL as the source of truth and SQLite as a disposable cache, git worktrees instead of containers for isolation, and no approval gates by default to maintain development velocity. Stoneforge supports integration with GitHub, Linear, Notion, and Obsidian, and works with Claude Code, OpenAI Codex, and OpenCode. The system allows developers to run multiple AI coding sessions across different subscription plans simultaneously, helping to work around rate limits.
King built Stoneforge after experiencing burnout from manually coordinating 3-5 Claude Code instances, constantly preventing work overlap and re-explaining architectural decisions across sessions. When workers hit context limits, they commit their progress with handoff notes, allowing fresh agents to continue with full context. The platform differs from Claude's experimental agent teams feature by offering persistent state storage, a web dashboard with live monitoring, automatic worktree isolation, and multi-provider support.
- Supports multiple AI providers (Claude Code, OpenAI Codex, OpenCode) and allows splitting work across different subscription plans to bypass rate limits
- Released under Apache 2.0 license with integrations for GitHub, Linear, Notion, and Obsidian
Editorial Opinion
Stoneforge tackles a real pain point that's emerged as AI coding assistants have matured — the ironically manual overhead of coordinating multiple AI agents. The architectural choices are particularly thoughtful: using git worktrees instead of containers is elegant, and the event-sourced design with JSONL provides both auditability and resilience. The decision to remove approval gates and push review to the merge level is bold but pragmatic. This feels like infrastructure that the AI-assisted development ecosystem genuinely needs as developers increasingly orchestrate multiple agents rather than writing code directly.
