Skip to main content
The Orchestrator is the central intelligence layer of the Corti Agentic Framework. It serves as the primary interface between users and the multi-agent system, coordinating the flow of conversations and tasks.
Diagram showing guardrails in the agentic framework

​
What the Orchestrator Does

The Orchestrator reasons about incoming requests and determines how to fulfill them by coordinating with specialized Experts. Its core responsibilities include:
  • Reasoning and planning: Analyzes user requests and determines the necessary steps to complete them
  • Expert selection: Decides which Expert(s) to call, in what order, and with what data
  • Task decomposition: Breaks complex requests into discrete tasks that can be handled by individual Experts
  • Response generation: Aggregates results from Experts and typically generates the final response to the user
  • Context management: Has full access to the context, while Experts typically only have scoped access to relevant portions
  • Safety enforcement: Enforces guardrails, type validation, and policy-driven constraints to ensure safe operation in clinical environments
The Orchestrator does not perform specialized work itself—instead, it delegates to appropriate Experts and coordinates their activities to accomplish complex workflows.

​
Safety Mechanisms

​
Max Iterations

The agent loop runs a maximum of 5 iterations by default (configurable per expert). After hitting the limit, it enters forceFinish mode, which makes up to 2 additional LLM calls restricted to terminal tools (complete_tool, input_required_tool, task_failure_tool). If those also fail to produce a terminal result, the task fails with “Max attempts reached without a finishing tool call.”

​
Credit Pre-flight Check

For orchestrator runs, the agent performs a pre-flight credit check before processing. If the metering service reports insufficient credits, the task is immediately rejected with state rejected and a message about insufficient credits.

​
Expert Invocation Progress

When the orchestrator calls an expert tool, it emits a Working status update with the text “Calling expert: <expert-name>”. These intermediate status messages are part of the task’s history but are not user-facing. If you are polling for task status, you may see these intermediate working states.
For more information about how the Orchestrator fits into the overall architecture, see Architecture. To understand how context and memory work, see Context & Memory.
Please contact us if you need more information about the Orchestrator in the Corti Agentic Framework.