Astra¶
Astra is an operating system for autonomous agents: a microkernel (actors, task graph, scheduler, messaging, state), sixteen microservices, sandboxed tools, layered memory, LLM routing, real-time chat, and platform stability infrastructure — not a single-model chat wrapper or a UI product.
Executive summary: Persistent agents submit goals; the planner materialises DAGs of tasks; the scheduler dispatches work via Redis Streams; workers run tasks inside sandboxes; Postgres is the source of truth and Redis/Memcached enforce the ≤10ms read path. Everything external talks through JWT; services talk over mTLS; dangerous work passes policy and approval gates. Agents can have profiles (system prompts, attached documents) that propagate through planning to execution, and users can interact via WebSocket chat or Slack.
flowchart LR
U[Client] --> GW[API gateway]
GW --> GS[goal-service]
GS --> PL[planner]
PL --> TS[task-service]
TS --> SCH[scheduler]
SCH --> R[(Redis)]
R --> EW[execution-worker]
EW --> TR[tool-runtime]
Goals¶
| Goal | Target |
|---|---|
| Scale | Millions of agents, 100M+ tasks/day (PRD §1). |
| Latency | No hot-path API read over 10ms p99; scheduling median ≤50ms, P95 ≤500ms (PRD §25). |
| Safety | Sandboxed tools, RBAC, approvals, secrets in Vault, mTLS everywhere between services. |
| Operability | Metrics, traces, runbooks, rolling upgrades with backward-compatible schema. |
| Resilience | Agent restore on startup, dead-letter tasks, circuit breakers, consumer retry, goal idempotency (PRD §21, P0-P2). |
Core capabilities (PRD v3.0)¶
- Platform stability (P0-P2): Agent restore on startup, task dead-letter queue, Redis consumer retry/reclaim, readiness vs liveness probes, gateway circuit breakers, goal idempotency, configurable task-stream sharding, supervisor wiring, mailbox-full handling.
- Agent profile & context (Phase 9): System prompts, attached documents (rules, skills, context docs, references), context propagation through planning and execution pipeline.
- Real-time chat agents (Phase 10): WebSocket streaming with tool invocation, session management, message injection.
- Slack integration (Phase 12): Connect chat agents to Slack workspaces, proactive posting, platform-configurable Slack app secrets.
- Hardware acceleration: Metal/Neural Engine on macOS, CUDA on Linux, graceful CPU fallback. macOS is a supported production target.
- Olympus application layer: External agent adapter framework, webhook ingest, goal-level dependencies, agent-to-agent goal posting, dual-approval, trust scores.
Non-goals¶
- Not building foundation models — Astra integrates providers.
- Not replacing every data platform — it composes Postgres, Redis, object storage, etc.
- Not embedding app logic in the kernel — strict kernel/SDK/app boundary (PRD §2).
- Not only a chat product — chat is one surface on the gateway; core is the task graph and actor runtime.
Who should read what¶
| Role | Start with |
|---|---|
| New contributor | Getting started → Glossary → Architecture overview |
| Backend engineer | Kernel, Task graph, Services |
| Operator | Operations, Deployment, runbooks |
| Security reviewer | Security, PRD §18 |
Reading order¶
- This page → Glossary if terms are unfamiliar.
- Architecture overview — layers and goal→task flow.
- Services — all sixteen services.
- Reference — contracts, schema, Redis, APIs.
- Operations when you’re on call.
At a glance¶
| Dimension | Value |
|---|---|
| Language | Go (primary), Python tooling |
| Kernel | Microkernel + actor runtime |
| Tasks | Distributed DAG, transactional state |
| Bus | Redis Streams + consumer groups |
| Source of truth | Postgres (+ pgvector) |
| Hot cache | Redis, Memcached |
| Sandboxing | WASM / Docker / Firecracker |
| Services | 16 canonical microservices |
| Spec | PRD in the Astra repo |
Maturity
Astra tracks Engineering Specification v3.0 (PRD). Phases 0–10 are complete; Phase 11 (multi-tenancy) is in progress; Phase 12 (Slack) is partial. When in doubt, read the PRD section cited on each page.
Sections¶
-
Getting Started
Repo layout, prerequisites, local paths.
-
Glossary
Terms and acronyms.
-
Security
mTLS, JWT, sandbox, Vault, approvals.
-
Architecture
Kernel, actors, scheduler, memory, LLM routing.
-
Reference
gRPC, schema, Redis, APIs, metrics, SLAs.
-
Operations
Runbooks and incident flow.
-
Deployment
K8s, local, macOS, GCP.