Warren

0Popularity

About Warren

Warren provides infrastructure for executing coding-agent workloads as isolated, observable jobs on infrastructure under user control. It manages the full lifecycle of agent runs from workspace preparation through Git delivery, including sandboxed execution, live event streaming, spend limits, and optional intervention. The system supports multiple execution backends such as local boxes with bwrap, sibling Docker containers, or Kubernetes pods, and includes runtime adapters for agents like Pi and Claude Code. Each run receives a disposable workspace, and Warren persists run records, outcomes, costs, and Git delivery facts for later inspection. Operators can start runs via CLI, schedule, or HTTP API, with concurrency and budget limits enforced during execution. The base service records structured events and outcomes, while optional extensions provide additional capabilities such as append-only audit logs or independent verdicts on finished runs.

Key features

Isolated sandboxed execution per run
Live event streaming and structured run records
Spend and concurrency limits enforcement
Workspace refresh and disposable run branches
Git delivery with optional pull request creation
Multiple execution backends (bwrap, Docker, Kubernetes)
Runtime adapters for Pi and Claude Code
Optional audit-log and judge extensions

Use cases

Running unattended agent workloads that must survive failures
Scheduling recurring agent tasks with budget enforcement
Sharing a single deployment among a trusted team for collaborative agent runs

Pros

Manages full agent-run lifecycle from workspace to Git delivery
Supports multiple execution backends (local, Docker, Kubernetes)
Persists structured run records for audit and review
Enforces concurrency and spend limits during execution
Includes runtime adapters for multiple coding agents

Cons

No named user or RBAC support in current release
Single bearer token guards the deployment
No per-user attribution in current release

Frequently asked questions about Warren

What is Warren and what does it do?

Warren provides infrastructure for executing coding-agent workloads as isolated, observable jobs on user-controlled infrastructure. It manages the full lifecycle of agent runs, including workspace preparation, sandboxed execution, live event streaming, spend limits, and Git delivery.

Who should use Warren?

Warren is useful when an agent run transitions from a terminal session to a workload that may need to continue unattended, repeat on a schedule, survive failure, or be visible to others. It suits teams or individuals who require durable evidence and infrastructure control.

How does Warren work?

Warren operates the run lifecycle from workspace preparation to Git delivery. It isolates each run in a disposable workspace, enforces limits, streams live events, and delivers changes via Git. Operators can start runs via CLI, schedule, or HTTP API.

What execution backends does Warren support?

Warren supports multiple execution backends, including local boxes with bwrap, sibling Docker containers, or Kubernetes pods. The backend is selected via the WARREN_RUNTIME environment variable.

Can Warren integrate with existing coding agents?

Yes, Warren includes runtime adapters for agents like Pi and Claude Code. Any harness with a Warren runtime adapter can use its run model and API.

How do I get started with Warren?

Start by setting up Warren on your infrastructure, selecting an execution backend, and configuring a runtime adapter for your agent. Use the CLI, schedule, or HTTP API to dispatch runs, then inspect the persisted run records for outcomes and events.