Concepts

The ideas that shape Arbiter: how agents speak to the runtime, how work is supervised, how memory persists, and how events become runs.

Start here if you want the model before the reference material.

Concept	What it covers
Writ	The slash-command DSL agents emit inline
Advisor	Structural supervision gates (CONTINUE / REDIRECT / HALT)
SSE events	The stream contract shared by TUI and HTTP
Structured memory	Typed, temporal facts with layered retrieval
Tenants	Isolation boundaries in API mode
Authentication	Bearer tokens and tenant scope
Scheduler	Natural-language schedules and run history
Sandbox	Constrained execution for risky tools
MCP	Stdio and remote tool servers
A2A	Agent-to-agent federation
Artifacts	Persisted files tied to conversations
Todos	Progress checklists agents and clients share
Lessons	Durable learnings extracted from runs
Search	External search fallbacks for /search
Data model	Core objects and field schemas
Durable execution	In-flight requests you can reconnect to
Fleet streaming	Multiplexed streams across agents
Operations	Deploy notes, proxies, and observability