Orchestrating agents

Edit online

Agent orchestration enables you to build scalable, multi‑agent systems where a primary agent coordinates specialized collaborator agents to complete complex workflows. Instead of relying on a single agent to perform every task, orchestration enables you to compose agents that each bring focused expertise, enabling accuracy, maintainability, and reuse across automation scenarios.

Before you begin

Before you build an orchestrated agent system, you must be a Builder or an Admin. These roles define how you contribute to the orchestration lifecycle.
To get started, make sure you understand your workflow, identify which steps require specialized agents, and confirm that you have access to the necessary app connections. This preparation helps you design effective delegation logic and create collaborator agents that work together smoothly.

What agent orchestration is

Agent orchestration is a pattern where a primary agent delegates work to one or more collaborator agents, each focused on a specific domain to complete tasks. Instead of building a single agent that tries to handle every step, you create a system of specialized agents that work together through structured delegation.

Orchestration enables the primary agent to:

Delegate tasks to collaborators with the right expertise
Compose results from multiple agents into a single response
Reuse collaborators across multiple workflows
Support nested collaboration, where collaborators can delegate to their own collaborators

The collaborator agents extend the primary agent's capabilities by contributing their artifacts (outputs, data, or results) for use during the primary agent's response generation.

Why orchestration works better than a single agent

Orchestration creates a more scalable and reliable system by distributing responsibilities:

Specialization: Each agent focuses on a specific domain, reducing errors and improving accuracy.
Maintainability: You can update or replace a collaborator without affecting other parts of the system.
Reusability: A single collaborator such as a Calendar Manager can support many primary agents and workflows.
Clear boundaries: Each agent has a defined role, making orchestration easier to understand, test, govern, and extend.

How orchestrated systems work

In an orchestrated system, the primary agent coordinates the entire workflow. It receives the user request, determines which collaborator agents to involve, delegates tasks in sequence, gathers each collaborator’s artifacts, and synthesizes the final response.

Receives the user request and analyzes what capabilities are needed
Evaluates handoff conditions to determine which collaborator agents should be involved
Delegates tasks to appropriate collaborators sequentially
Collects artifacts (outputs) returned by each collaborator
Generate the final response using the gathered artifacts and its own logic

Basic orchestration flow: The primary agent evaluates the request, delegates to collaborators sequentially, and synthesizes their artifacts into a final response:


User Request
     ↓
Primary Agent
     ↓
[Evaluates handoff conditions]
     ↓
Collaborator Agent A → Artifact A
     ↓
Collaborator Agent B → Artifact B
     ↓
Primary Agent
[Synthesizes response using Artifacts A & B]
     ↓
Final Response to User

Nested collaboration: Collaborator agents can have their own collaborators, creating an interconnected network. The primary agent can delegate through multiple levels:


Primary Agent
     ↓
Collaborator Agent A
     ↓
Collaborator Agent A1 → Artifact
     ↓
Collaborator Agent A2 → Artifact
     ↓
[Results flow back to Primary Agent]

This delegation chain allows for sophisticated task decomposition while maintaining clear coordination paths.

Sequential execution model

When a primary agent delegates to multiple collaborators, each one completes its work fully before the next begins. This sequential processing is the standard execution model and there is no configuration to enable parallel execution.

How sequential execution works

The first collaborator receives and completes its task
The next collaborator starts only after the previous one finishes
Each collaborator can use artifacts created earlier
The primary agent composes the final response after all collaborators finish

Important:

The collaborator agents in watsonx Orchestrate run sequentially. Currently, watsonx Orchestrate does not support parallel execution of collaborator agents or tools, and there is no configuration available to enable parallel processing from the Building agents and tools page.

However, parallel execution of agents or tools is supported through Agentic workflows where Agent Nodes can be added in parallel branches to run concurrently.

Parallel processing is available through:

Pro‑code agentic workflows (ADK) using Parallel For each and Parallel Branch.

Adding a collaborator agent

On the agent configuration page, click Toolset > Add agent.
Choose from where you want to add the agent:

To remove a collaborator agent, go to the Toolset, locate the agent you want to remove, click the vertical ellipsis next to it, and select Remove.

Choosing the right method


Method	When to use	Sample use case
From catalog	When you want ready-to-use or partner-provided agents	Adding a prebuilt HR or IT support agent
From local instance	When your team already created agents in the same watsonx Orchestrate instance	Reusing an internal finance or compliance agent
From external sources	When you want to connect third party, watsonx.ai, or Assistant agents	Using an AI model that is deployed on another platform
Using ADK	When your team develops external A2A agents that must work with native ones	Adding a custom logic agent that is built by using Orchestrate ADK

Note:

You can build and register external agents by using the Agent Development Kit (ADK). For more information about how to build, configure, and register agents with ADK, see ADK documentation.

What to do next

After you understand how agent orchestration works, you can continue building your agent by defining collaborators, configuring handoff rules, and validating your end‑to‑end workflow.

Start by identifying which steps require specialized agents, create those collaborators, and map the data each collaborator needs and returns. Then test your orchestration with real scenarios to ensure agents delegate correctly, share artifacts as expected, and produce clear, accurate responses.