@claude Research and evaluate sandbox environments for self-hosted AI agents

[Kavinjsir]

Context

There is growing demand for Cloud AI Agent sandboxed execution environments. Platforms like e2b.dev (used by Perplexity Computer and Manus) and modal.com sandboxes offer managed solutions, and we have startup credits for Modal. The team wants to evaluate these options and potentially build our own solution—possibly by modifying Ray internals or leveraging Temporal/LangGraph on one of our subnets—to provide the most intuitive and efficient infra for agent sandbox execution.

What needs to happen

1. Research existing platforms: Evaluate e2b.dev and Modal sandboxes for capabilities, pricing, security, and integration ergonomics.
2. Prototype with existing platforms: Try e2b.dev and/or Modal sandboxes with our self-hosted agent setup to understand developer experience and limitations.
3. Assess build-vs-buy: Determine whether an existing platform meets our needs or if we should build a custom sandbox solution.
4. Explore custom sandbox architecture: If building our own, investigate:
   • Modifying Ray internals (building on existing Ray/KubeRay infrastructure)
   • Temporal or LangGraph-based orchestration on a Poseidon subnet
5. Document findings and recommendation: Produce a short write-up comparing options with a recommended path forward, including estimated effort and subnet integration opportunities.

Acceptance criteria

• A comparison document covering e2b.dev, Modal sandboxes, and custom-build approaches (pros/cons, cost, security, scalability)
• At least one working prototype demonstrating agent code execution in a sandboxed environment
• Clear recommendation on which approach to pursue for self-hosted agents
• If a subnet opportunity is identified, outline a high-level design for how sandbox infra could be offered as a subnet service