Shared control plane
One mvsqlite-backed SQLite namespace coordinates tenants, nodes, schedules, claims, leases, and heartbeats.
Architecture
A distributed runtime with SQLite-shaped state.
Willow separates control-plane coordination from per-agent durable state. mvsqlite and mvstore provide SQLite semantics over FoundationDB; NATS is used for coordination and handoff, not as the system of record.
Storage architecture
mvsqlite + FoundationDB for durable agent state
zero-copy forks
willow_control
Per-agent namespaces
Authenticated namespaces hold each agent’s VFS, messages, sessions, waits, tool calls, and checkpoints.
Distributed storage substrate
mvstore and mvsqlite provide SQLite semantics over FoundationDB, with MVCC snapshots and overlay forks.
01
Serializable claims
Identical Willow nodes claim ready agents through serializable mvsqlite transactions and renew worker leases while they run.
02
Tenant isolation
The control plane separates tenants and auth metadata while every agent persists into its own authenticated namespace.
03
Overlay forks
Forking an agent is metadata-only: the new namespace overlays a source namespace at a specific mvsqlite version.