effect

Effect types for the sans-IO agent core.

Effects are declarations of intent — the core describes what it wants done, and a runtime interprets these effects against the real world.

Two effects cover the entire agent loop: CallProvider — send messages to the LLM ExecuteTools — run tool calls

The on_response / on_results callbacks wrap results into the core’s message type, closing the loop without the core knowing how the effects were executed.

Types

Effect

</>

An effect request from the pure core. The runtime interprets these.

pub type Effect(msg) {
  CallProvider(
    messages: List(message.Message),
    tools: List(tool_definition.ToolDefinition),
    on_response: fn(
      Result(provider.InferenceResult, error.AiError),
    ) -> msg,
  )
  ExecuteTools(
    calls: List(message.ToolCall),
    on_results: fn(
      List(#(message.ToolCall, Result(json.Json, tool.ToolError))),
    ) -> msg,
  )
}

Constructors

CallProvider(
  messages: List(message.Message),
  tools: List(tool_definition.ToolDefinition),
  on_response: fn(Result(provider.InferenceResult, error.AiError)) -> msg,
)

Send messages and tool definitions to the LLM provider. The runtime calls the provider function, then feeds the response back via the on_response callback.

ExecuteTools(
  calls: List(message.ToolCall),
  on_results: fn(
    List(#(message.ToolCall, Result(json.Json, tool.ToolError))),
  ) -> msg,
)

Execute a list of tool calls. The runtime runs each tool (possibly in parallel), then feeds all results back via the on_results callback.