Architecture¶
This section describes how BOS is built and the principles that guide its design. It is the up-to-date source of truth — for the deeper, point-in-time rationale behind individual decisions, see the BEP Index.
Design principles¶
- From zero to agent in one command. Sensible defaults; complexity is opt-in. A bare config runs; every layer (channels, sub-agents, custom tools) is something you add.
- Minimal to start, project-ready. Start minimal and grow into a multi-agent project by editing one TOML file and dropping files into conventional directories — no rewrites.
- Micro-kernel and plugins. A small core with well-defined extension points rather than a monolith. Tools, providers, channels, stores, consolidators, and plugins are all named, interchangeable implementations registered at an extension point.
- Clean architecture. Dependencies point inward through concentric rings; the agent core stays free of framework and I/O concerns. Outer rings (harness, gateway, CLI) depend inward on the core, never the reverse.
How the pieces fit¶
┌─────────────────────────── gateway process ───────────────────────────┐
external client │ channel ──▶ mailbox ──▶ actor ──▶ agent (LLM loop) ──▶ tools/plugins │
(TUI/Telegram) │ ▲ │ │ │
│ └──────── reply ─────────┘ harness services: │
│ chat_store, consolidator, │
│ mail_route, job_runner │
└────────────────────────────────────────────────────────────────────────┘
- Gateway — the process that hosts actors and channels and serves a control plane. See Runtime & gateway.
- Actors & agents — named, addressable runtime instances bound to LLM-driven agents. See Agents & actors.
- Channels — how users reach an agent (TUI, Telegram, Lark, HTTP). See Writing channels.
- Harness — the lifecycle owner of shared services (chat persistence, memory
consolidation, message routing, background jobs), each selected by name in
[harness]. - Extension points — the micro-kernel seams. See Extending BOS.
- Memory & skills — platform-managed memory with off-turn consolidation, and progressively-disclosed skill playbooks. See Memory & skills.
Where to read more¶
- Concepts — the mental models in depth.
- Configuration — how each part is configured.
- Extending BOS — the extension-point model and how to build on it.
llm-full.md— a single dense, code-grounded reference of every mechanism, intended for ingestion by an AI agent.boscli initdrops a copy into every project.
Deeper detail
The BEP Index links to the enhancement proposals that record the full design reasoning. BEPs capture point-in-time decisions and may contain detail that has since evolved — this section is the current reference.