OpenClaw Creator's new secret project...

Key Concepts

• Crabbox: A tool designed to manage isolated, cloud-based development environments for AI agents, enabling parallel testing without local resource conflicts.
• Codebase Harness: A framework providing agents with the necessary tools and skills to verify their own work and generate evidence.
• Work Tree: An isolated directory structure for managing different branches or tasks simultaneously.
• Sandbox Isolation: The practice of running development servers in separate, ephemeral cloud containers to prevent database or port conflicts.
• Artifacts: Evidence generated by agents (screenshots, video recordings, logs) to prove code functionality within a Pull Request (PR).
• Syncing (Dirty Diffs): The process of pushing uncommitted local changes to a cloud sandbox for real-time testing without requiring a formal Git commit.

---

1. The Problem: Scaling AI Agent Workflows

The speaker highlights that the bottleneck in modern AI-assisted development is no longer writing code, but the review and verification process. When running multiple agents in parallel (10–15+ sessions), several issues arise:

• Resource Contention: Local machines struggle with CPU/RAM usage when running multiple Docker daemons or databases (e.g., Supabase).
• Port Conflicts: Hard-coded ports in applications prevent multiple instances from running locally.
• State Corruption: Multiple agents modifying the same database schema simultaneously lead to system instability.
• CI/CD Overhead: Traditional commit-push-CI flows are too slow and create "noisy" commit histories when agents are debugging.

2. The Solution: Crabbox Framework

Crabbox allows agents to spin up ephemeral, isolated cloud environments ("boxes") to test code in real-time.

Step-by-Step Implementation Process:

1. Define Environment: Create a Dockerfile that encapsulates all dependencies (Node, CLI tools, etc.).
2. Configure crabbox.yml: Define the sandbox provider (e.g., Daytona), snapshot images, excluded files (to optimize sync speed), and environment variables.
3. Setup Script (setup.sh): A script that automates the installation of dependencies and the launching of the development server.
4. Warm-up & Sync:
   • crabbox warm-up: Provisions the cloud sandbox.
   • crabbox run: Syncs uncommitted "dirty" changes from the local machine to the cloud and executes bash commands.
5. Testing & Evidence: Agents use tools like Playwright to perform end-to-end tests.
6. Artifact Collection: Use artifact publish to upload screenshots or videos to S3 buckets or GitHub release assets, which are then linked in the PR.
7. Cleanup: crabbox stop terminates and deletes the sandbox to save resources.

3. Key Arguments and Perspectives

• Evidence-Based Merging: The speaker argues that agents must provide "proof" (artifacts) to make PRs reviewable. Without this, the burden on human reviewers becomes unsustainable.
• Local vs. Cloud: While local development is fine for 1–2 agents, scaling requires moving the "dev server" to the cloud. The speaker emphasizes that the local machine should only be used for editing, while the cloud handles the execution and verification.
• Efficiency: By syncing only the "dirty diffs" (uncommitted changes), developers can re-test fixes in seconds rather than waiting for full CI pipeline rebuilds.

4. Technical Details & Tools

• Daytona: Used as the sandbox provider for its fast startup times.
• SSH Tunneling: The speaker demonstrates how to tunnel local ports to the cloud sandbox, though notes that for agent-driven testing, it is better to have the agent run the tests directly in the cloud to avoid local port conflicts.
• cbx.sh Wrapper: A custom bash script recommended to handle the asynchronous nature of cloud provisioning (e.g., polling for status until the server is ready).

5. Notable Quotes

• "The bottleneck is no longer write code, but actually how do you get code merged into the code base."
• "If you're running a local Supabase... all those environments are actually sharing the same instance. If one of them is trying out a new schema, they can very easily break all the other sessions."

6. Synthesis/Conclusion

Crabbox addresses the critical scaling challenge of AI-driven development by decoupling the editing environment (local) from the execution/testing environment (cloud). By providing agents with a standardized "harness" to spin up isolated sandboxes, sync uncommitted changes, and generate visual evidence, teams can significantly increase their PR throughput while maintaining system stability. The framework is open-source and designed to be integrated directly into an agent's skill set, allowing for autonomous, end-to-end verification of code changes.