Skip to main content
Orchestration determines how the agent decides which capability (Action or Subagent) to use for a given request. Navigate to the Configuration tab in the Portal to manage these settings. Orchestration

Execution Modes

When you call an agent, the execution_mode parameter controls how the agent processes the request:

route — Single-Step Routing

The agent selects one capability and returns the result. Best for simple, direct requests where a single tool call is sufficient.

reasoning — Multi-Step Orchestration

The agent enters a reasoning loop powered by Cortex, Maitai’s event-driven orchestration engine:
  1. Plan — An LLM orchestrator reviews the request and available capabilities, then plans which actions to run.
  2. Execute — The planned actions fire and produce results.
  3. Process — Results are summarized and progress is evaluated.
  4. Repeat or respond — The orchestrator decides whether more work is needed or the agent is ready to respond.
This mode handles complex, multi-step tasks that require gathering data from multiple sources, delegating to sub-agents, and synthesizing results.

auto — Automatic Selection (Default)

The agent picks the best mode automatically. It selects reasoning if sub-agents exist or multiple actions are available; otherwise falls back to route.

Routing Strategy

Within orchestration, you can configure how routing decisions are made:
  • Flexible (LLM): The agent uses an LLM to intelligently select the best tool based on the user’s intent, the agent’s state, and the available capabilities.
  • Deterministic (Rules): Use specific logic-based rules to route requests to specific targets. This is ideal for high-precision workflows where behavior must be strictly controlled.

Orchestration Instructions

Provide additional guidance that the LLM uses specifically during the routing/planning phase. These instructions help the agent weigh different capabilities and make the right decision when user intent is ambiguous.

Performance & Chaining Options

These optional flags can be passed at request time to tune agent behavior. See Agent Call for the full parameter reference.
ParameterDefaultDescription
high_performancefalseFlattens sub-agent actions into the root agent, running everything in-process. Eliminates inter-process overhead for agents with sub-agent hierarchies.
disable_chainingfalseWhen false, the orchestrator can plan multi-action chains in a single call (e.g., “fetch data, then enrich it”). When true, plans one action at a time for simpler, more predictable execution.
disable_strategyfalseWhen true, skips the strategy-building step and goes straight to action planning.

Next