01
Safe Multi-Agent Parallelism
Independent agent sandboxes with controlled lifecycle boundaries.
Infrastructure for parallel coding agents
A safe sandbox for every agent.
Branch once. Run many agents. Publish safely.
Give every coding agent an isolated, delta-efficient workspace. Execute concurrently, trace published changes, and reconcile results without last-writer-wins data loss.
- COW Workspaces
- CLI + MCP
- Linux AMD64 + ARM64
How parallel agents work
One base. Many sandboxes. One safe merge.
Every agent branches from a shared immutable base, works in isolation, and converges at a publication gate that only accepts fully resolved changes.
Immutable base workspace
shared source of truth
Publication gate cleanconflict
new immutable layer published
01
Branch from one base
A single immutable base workspace is the shared source of truth for every agent.
02
Run agents in parallel
Each agent gets its own wooden sandbox with a private copy-on-write delta layer.
03
Reconcile at the gate
Branches converge at a publication gate. Clean changes pass; conflicts stop safely.
One connected system
Six capabilities, one coherent architecture.
Each capability builds on the same isolated, copy-on-write foundation — from spawning agents to publishing their work atomically.
02
Delta-Efficient COW Workspaces
One shared immutable base with private storage for metadata and changed bytes.
03
Published-Change Provenance
Attribute published lines to the responsible workspace session or operation.
04
Agent-Native Tool Plane
Shell, yield, stdin, cancellation, read, write, edit, list, and blame through CLI and MCP.
05
Low-Friction Linux Deployment
Inject the sandbox runtime into compatible Linux container images without maintaining a custom image.
06
Conflict-Aware Publication
Detect concurrent changes, merge eligible text edits, and publish the complete changeset atomically.
Featured — conflict-aware publication
All resolved, or nothing published.
Publication is transactional. Eligible text edits are merged three-way against the base. If concurrent changes conflict, the entire changeset is rejected — no partial writes, no last-writer-wins data loss.
Clean merge — published as a new immutable layer
Conflict detected — no partial changes published
Agent Base
export function auth() {return session}
Current Workspace
export function auth() {validate(session)return session}
Agent Changes
export function auth() {log(session.id)return session}
Publication gate · atomic
Developer experience
One control plane, from shell to provenance.
Drive workspaces through a CLI or wire them into agents over MCP. The same tool plane covers files, provenance, and live observability.
terminal
# spawn an isolated sandbox from the base$ sandbox create --base main --name agent-aworkspace agent-a ready (cow delta attached)$ sandbox run agent-a -- "apply refactor"running in isolated workspace…
Give every agent room to work—without losing control of the result.
Branch once, run agents in parallel, and publish atomically on a shared immutable base.