Today, Fulcrum is announcing Druids.
Druids is a batteries-included library to coordinate and deploy coding agents across machines. Our beta users have used Druids to work on open math problems, conduct ML “autoresearch”, and make software orders of magnitude faster.
The Motivation: Automate Your Research and Engineering Processes
Research and software development advance through repeated cycles of search, evaluation, and refinement. These cycles are bottlenecked by the cost of human cognitive labor, limiting both how many cycles you can run and how broadly you can search.
LLMs have made individual steps in these cycles faster, but both research and engineering involves steps that depend on each other, branch conditionally, and benefit from running in parallel. A single agent can’t orchestrate the entire process effectively. To automate these cycles, you need a way to coordinate many agents into a coherent process, with rules for interaction and systems for verification.
Druids provides the language and infrastructure for this coordination. Users can write Python programs that specify agents, their goals, and how they talk to each other. Druids provisions full VMs so that these agents can run reliably and communicate effectively.
The Language: Easily Write Reliable State Machines
Druids programs are state machines. You define what your program tracks, the events that move it forward, and how agents handle errors and branching. Because every dependency is explicit, Druids makes it easy to parallelize your agents.
Events allow you to inject deterministic structure and control flow to structure how your agents work towards the task of the program. They are useful for defining controlled steps and flows, like:
- forcing a model to iterate against hard tests and harness signals
- building a verification hierarchy, where agents spawn outputs that are verified and redteamed by other agents until they match a set of properties
- controlling distributed task state, like having a lock around the ways agents write to shared resources or user-facing systems
This means your agent software is predictable, reproducible, and testable. You can focus on specifying the shape of your agent team, while Druids handles the infrastructure.
The Infrastructure: Pre-warmed Clouds Made for Agents
Each agent gets a full sandboxed VM with your repo cloned and dependencies installed. This means that agents have the permissions they need to work freely, including full shell access, package installation, and file system writes. They can actually verify their own work by running code, executing tests, and checking results before moving on. And because runs execute on real infrastructure, they can go for hours or days without disconnecting. You can choose to steer the process at any point by inspecting the program state and messaging the agents.
Druids ships with a UI to visualize and steer your agent team, a CLI to launch and manage runs from your terminal, and an mcp so your local coding agent can call Druids.
Try it today
We believe Druids is the right way to build agent software. We’re open-sourcing it today and we’d love to hear what you think.
Links: