High-level overview of our main strategic priorities for 2026:
- Overhaul Documentation - Restructure and expand current documentation to lower the barrier to entry for new users.
- Website Refresh [#166] - Update the website to accurately reflect the latest features, documentation links, and usage examples.
- PyPI Distribution [#146] - Publish the agent-sandbox-client package to pip for easier installation
- Expand SDK functionality - natively support methods like read, write, run_code, etc. within the Python SDK
- Benchmarking Guide
- Strict Sandbox-to-Pod Mapping [#127] - Ensure a reliable 1-to-1 mapping exists between a Sandbox and a Pod
- Expand Sandbox use cases - Computer use case, browser use case, and base images
- Decouple API from Runtime - enable full customization of runtime environment without breaking API
- Implement GO Client [#227]
- Scale-down / Resume PVC based - Pause resume preserving PVC only, when replicas scale to 0, PVC is saved, when sandbox scales back PVC is restored
- Add complete CR, SDK and template support
- API Support for Multi-Sandbox per Pod - Extend API to support multiple sandboxes in a Pod
- Startup Actions [#58] - Allow users to specify actions at startup, like immediately pausing the sandbox or pausing it at a specific time
- Auto-deletion of (bursty) sandboxes (RL training typical usage)
- Status Updates [#119] - Functionality to properly update and reflect the status of the sandbox
- Creation Latency Metrics [#123] - Add a custom metric to specifically track the latency of Sandbox creation time
- Runtime API OTEL/Tracing Instrumentation - Instrument runtime API with OpenTelemetry and Tracing to provide guidance on further instrumentation
- Metadata Propagation [#174] - Ensure that labels and annotations are correctly propagated to sandbox pods
- Headless Service Port Handling [#154] - Ensure Headless Services correctly set ports when containerPort is configured
- Detailed logs Falco configuration extension - Propagate Falco configuration for gVisor sandbox logging. Enable configuration via Agent Sandbox API
- API Support for other isolation technologies - Continue extending the support to QEMU, Firecracker and other technologies; Process isolation (pydantic)
- OpenEnv Support [#132] - Develop support for AgentSandbox within the OpenEnv environment
- Agent & RL Framework Support - Tighter integration between Agent Sandbox & popular Agent & RL frameworks like CrewAI, Ray Rllib
- Integration with kAgent
- Integration with other Sandbox offerings
- Deliver Beta/GA versions