AgentConfig

Configuration for the LLM agent loop. Passed to ctx.useAgent().

Source: @durable-streams/darix-runtime

interface AgentConfig {
  systemPrompt: string
  model: string
  tools: AgentTool[]
  streamFn?: StreamFn
  testResponses?: string[] | TestResponseFn
}

Fields

Field	Type	Required	Description
systemPrompt	string	Yes	System prompt sent to the LLM on each step.
model	string	Yes	Model identifier (e.g. "claude-sonnet-4-5-20250929").
tools	AgentTool[]	Yes	Tools available to the LLM. Always include ctx.darixTools first. See AgentTool.
streamFn	StreamFn	No	Custom streaming function. Overrides the default LLM provider.
testResponses	string[] | TestResponseFn	No	Mock LLM responses for testing. When set, no real LLM calls are made.

TestResponseFn

type TestResponseFn = (
  message: string,
  bridge: OutboundBridgeHandle
) => Promise<string | undefined>

A function that receives the current conversation and returns a mock response string, or undefined to end the agent loop.

AgentHandle

Returned by ctx.useAgent(). Also available as ctx.agent.

interface AgentHandle {
  run(input?: string): Promise<AgentRunResult>
}

Method	Return Type	Description
run(input?)	Promise<AgentRunResult>	Execute the agent loop. Runs until the LLM stops or all tool calls complete.

Parameters:

Parameter	Type	Required	Description
input	string	No	Optional user message appended to the conversation before the agent loop starts.

AgentRunResult

interface AgentRunResult {
  result?: unknown
  writes: ChangeEvent[]
  toolCalls: Array<{ name: string; args: unknown; result: unknown }>
  usage: { tokens: number; duration: number }
}

Field	Type	Description
result	unknown	Final result value, if any.
writes	ChangeEvent[]	All events written to the entity stream during this run.
toolCalls	Array<{ name, args, result }>	Record of all tool calls made during this run.
usage	{ tokens: number; duration: number }	Token count and wall-clock duration in milliseconds.