Fulcrum Research released Druids, an open-source library that coordinates multiple coding agents across sandboxed VMs without making you manage the infrastructure. You write an async Python function describing how agents should collaborate. Druids spins up Docker containers, provisions agents, and routes messages between them.
The framework uses an event-driven model. Agents decide when to trigger events. Your program decides what happens next. Each agent gets its own VM with your repo and dependencies. Agents can share machines, transfer files, and work on separate git branches. On the hosted version at druids.dev, agents can fork themselves with instant copy-on-write clones.
Example programs show practical patterns. One runs a builder, critic, and auditor that iterate until all three agree. Another spawns finder agents that scan for tasks and create implementor-reviewer pairs dynamically. A third pits Claude against Codex on the same specification to see which model produces better code.
To run Druids locally you need Docker, the uv package manager, and an Anthropic API key. The architecture includes a FastAPI server, a CLI, a Python client library, a runtime SDK, and a Vue 3 dashboard for monitoring.
Druids enters a crowded field. ctx unifies Claude Code, Cursor, Codex in one workspace, though neither isolates agents in separate VMs by default. That isolation matters when agents are running pentests or modifying code autonomously. The tradeoff is overhead. Each VM adds startup time and eats resources.
The project is new and it shows. Documentation is sparse. The Anthropic API key requirement suggests limited model choice for now, though the Claude-versus-Codex example hints at broader provider support. The Docker-plus-uv dependency stack won't appeal to everyone. Fulcrum offers both the open-source release on GitHub and a commercial hosted service. Whether Druids catches on depends on whether VM-level isolation proves necessary or just adds friction.