Skip to content
Documentation
Using the cockpit
Orchestrator

Agent Orchestrator

Agent Orchestrator (/orchestrator) hands a task to one agent. The agent works through it in a loop: it thinks, acts, then observes the result. You watch each iteration arrive. To open it, go to Fetch and select Agent Orchestrator in the link row above the execution history.

### Choose the agent Pick an id from the **Agent** list, for example `twins`, `claims`, `compliance` or `reverse_logistics`. The default is `twins`. ### Describe the task Type it in **Command** (**Describe the task for the agent…**). ### Run Select **Run**. The button becomes **Stop** while the loop runs. Select **Stop** to stop following the run.

Reading the loop

Each iteration appears as a numbered card with Thought, Action and Observation. Its status is a word as well as a colour:

  • Thinking
  • Executing
  • Observing
  • Done
  • Failed
  • Iteration limit reached
  • Not started

When the loop ends in Done, Final result shows the agent’s answer. A run stops after eight iterations at most. Before the first run the page reads Issue a command to start the agent loop. A failed request shows the server’s message.

Each proposed action in the agent’s final answer is judged before it is accepted. See the judgement notes on Fetch.

Behind this page: RalphOrchestratorService.ExecuteLoop (server stream). See the API reference.

OCRConnections