Architecture¶
Astra is a layered system: a small kernel, services, and infrastructure. Start with Overview for diagrams and end-to-end flow.
Doc map¶
| Page | Question |
|---|---|
| Overview | Layers? Goal → task flow? |
| Kernel | What belongs in the kernel? |
| Actor framework | Mailboxes, supervision? |
| Task graph | DAG, task states? |
| Scheduler | Dispatch, shards, latency? |
| Services | The 16 services? |
| Memory | Memory tiers? |
| LLM routing | Models, cost, cache? |
| Multi-tenancy | Isolation, roles? |
Design principles (summary)¶
- Kernel stays small — product logic lives in services and the SDK.
- Database is source of truth for durable agent/task state; caches accelerate reads.
- Hot-path reads target strict latency goals (PRD §25).
- Writes are durable first, then events, then cache refresh.
- Audit trail supports recovery and compliance.
- Actors favour clear message boundaries over shared mutable state.
Note
Implementation specifics are intentionally not on this public wiki.