Architecture

Switchboard receives text from a channel, checks the request, chooses an agent and starts a run. The run calls a model and tools; the answer returns through the channel.

The four parts around the dispatcher

A channel (Slack, CLI, HTTP or MCP) moves messages. An agent is a work definition. A provider supplies the model. An executor supplies a place for tools to run. The dispatcher connects them and enforces routing and access rules.

Where work and state live

The bot process handles requests and holds provider credentials. The state Worker stores durable run, conversation and work records when configured. Repository work can use a resident checkout or a sandbox; local development can run tools on the bot host. Worker topology gives the deployment detail.

For the human view of the system, start with Data model, then How a request flows. Execution and trust explains why tools run behind a separate boundary.