Obelisk 0.42: Durable Agents, Layered Sandboxes

Obelisk 0.42 introduces enhancements for durable agents and layered sandboxes. Workflow progress is stored in a database, allowing deterministic replay from recorded activity results. Agents can now run on native V8, with Linux VM activities available for tools requiring them. The workflow-agent prototype can deploy, test, and fix applications on another Obelisk instance.

Policy is split between server.toml and app.toml, with the effective permission being the intersection. App.toml can request less than app.toml grants, and app.toml cannot switch on exec activities unless server.toml permits it. Agents can rewrite code and deployment.toml within granted permissions. A request for broader access requires a change to app.toml, providing a clear policy diff.

Secrets default to placeholders, replaced at the network edge, and can be exposed with exposed_secrets, requiring a grant in app.toml bound to a digest of the component and its secrets. Exec activities, now requiring approval from both platform and app admins, run processes outside the sandbox. The durable coding-agent prototype has separate JavaScript and Rust implementations.

Metrics show significant improvements in replay time, with V8 being about 25 times faster than Boa WASM for a 726-event conversation. The VM activity layer allows agentic workloads to run inside Linux VMs, with guest HTTP going through the same policy as other components.