Overview
Astrid explores how WebAssembly capsules can compose without inheriting authority: manifests and event routes lead to effect-specific gates, principal-scoped state, accountable operations, and a proposed native kernel. Its hosted runtime is concrete; native and verified-computation extensions remain evidence-gated.
Learning outcomes
- Astrid keeps model behavior, principal identity, permission, state, and audit evidence separate.
- WebAssembly sandboxing removes ambient access, but effect-specific gates still decide each host operation.
- Shared capsules can serve multiple principals when invocation-scoped state, routes, quotas, and logs remain separated.
- The hosted runtime is concrete; the native kernel and verified-computation plans remain conditional on executable evidence.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What problem does Astrid solve for agent systems?
Astrid governs how composable software causes effects without inheriting the host’s authority. Its runtime separates behavior, identity, permission, state, and evidence.
How does Astrid prevent a capsule from using the host’s permissions?
Astrid starts WebAssembly capsules without ambient host access and checks each external effect through typed, resource-specific gates. Manifests request scope but do not grant it.
How does Astrid isolate multiple principals sharing one capsule?
Astrid installs per-invocation principal overlays for state, secrets, quotas, cancellation, and logs while routing events by principal. Shared code does not imply shared authority.
Is Astrid already a native operating system?
No. Astrid’s current product is a hosted macOS and Linux runtime; its ring-zero capability microkernel is a proposed direction whose claims remain evidence-gated.
What does Astrid’s audit trail prove?
Astrid’s signed, hash-linked audit can record authorization proofs and outcomes, but it cannot prove that the model’s judgment was correct.