Architecture

For the full architecture document, see Architecture Overview.

Design Principles

1. Modular — Every component is independent and replaceable
2. Event-driven — All operations emit structured events
3. Framework-agnostic — No opinions about LLMs or agent frameworks
4. Async-first — All I/O is non-blocking
5. Type-safe — Full TypeScript with strict mode

Component Map

Kernel (orchestrator)
├── Process Manager — spawn, pause, resume, terminate
├── MemoryBus — pluggable backends (InMemory, SQLite)
├── ToolRouter — registry, permissions, rate limits
├── MessageBroker — pub/sub, direct, priority queues
├── Scheduler — sequential, parallel, race, conditional
├── LLM Agent — Anthropic + OpenAI with tool loops
├── MCP Client — stdio + SSE transport
├── Contracts — runtime input/output validation
├── Checkpointing — serialize/restore process state
└── Config — YAML config loader

Data Flow

kernel.execute(processId, { task: 'work' })
  │
  ├── validate input (contract)
  ├── build ExecutionContext (memory, tools, messaging)
  ├── call handler(ctx)
  │     ├── ctx.memory.get/set (→ MemoryBus → Backend)
  │     ├── ctx.useTool (→ ToolRouter → handler)
  │     ├── ctx.publish (→ MessageBroker → subscribers)
  │     └── return output
  ├── validate output (contract)
  ├── check SLA (latency)
  ├── read __llm_usage (token tracking)
  └── return ExecutionResult