Agentic workflows
Agentic workflows, previously known as flows, are specialized tools that allow an agent to run a sequence of activities withing a single, reusable structure. These activities include calling tools, prompting for user input, logical logic blocks, or branching logic.
Rather than handling each step individually, agents can start an angetic workflow to manage the entire process from beginning to end. Agentic workflows are ideal for tasks that require coordination across systems or multiple decision points.
For example, an agentic workflow can be created to handle customer onboarding: collecting information, creating accounts, sending welcome emails, and notifying internal teams. Once built, this agentic workflow can be reused across departments, triggered by agents whenever a new customer joins—no need to manually coordinate each step.
By using agentic workflows, business users gain:
-
Confidence that tasks are completed correctly and consistently.
-
Speed through automation of repetitive steps.
-
Visibility into how processes run and where bottlenecks occur.
-
Scalability to apply the same logic across teams, regions, or products.
Agentic Workflows in the context of workspaces in IBM Cloud
In watsonx Orchestrate on IBM Cloud environment only, workspaces organize and control access to workflows and tools during the build and development phase. Workspaces introduce isolation, governance, and role-based access control without requiring separate service instances for different departments or teams.
Workspaces do not affect the runtime execution of workflows in chat. Also, workspaces do not affect API-based workflow invocation, where all workflows are visible to all users.
Workspace access is governed by a combination of service instance roles and workspace-level permissions. Two types of workspaces are available:
-
Global workspace: Holds agentic workflows and tools that are not attached to any private workspace. Assets in the global workspace remain visible to all builders and support shared, cross-team use.
-
Private workspace: Allows a specific set of builders to work on agentic workflows without making those assets visible to other builders. Only explicitly added users can access and modify artifacts in the private workspace.
For more information about workspaces, see Overview of workspaces.
Copying agentic workflows
Tools and workflows cannot be copied directly to another workspace. However, when an agent is copied between workspaces, its associated workflows and tools are copied with it. For more information about copying agents, see Managing workspace artifacts .
Balancing probabilistic agents with deterministic workflows
Agents use a probabilistic method to complete tasks. They sequence tools by using large language models (LLMs) and AI algorithms that follow natural language instructions. This approach gives you a flexible way to get work done by asking the agent to perform tasks through a chat or voice channel.
However, a fully probabilistic approach can sometimes be unpredictable, especially for complex processes that involve many tools or multiple agents. You can reduce this unpredictability by enabling agents to use agentic workflows as tools. Agentic workflows provide deterministic sequencing and support features such as loops, branching logic, state management, and parallel processing.
A purely deterministic approach can also be restrictive. By combining agents with agentic workflows, builders can balance probabilistic and deterministic behavior to deliver a flexible and predictable experience for users.
About agent patterns
The field of AI agents and multi-agent orchestration is developing quickly, with new patterns and techniques emerging regularly. Common agent patterns include:
-
Linear sequence (pipeline)
-
Supervisor
-
Multi-level supervisor
-
Vote and dispatch
-
Mixture of experts
-
Crowdsourcing
-
Swarm
You can implement these patterns by using agents or agentic workflows. Use agent instructions when you need flexibility and the pattern is simple. For patterns that require higher speed, reliability, or more complex behavior such as loops, branching logic, state management, or parallel processing, agentic workflows are a better choice. Agent nodes within agentic workflows make these patterns possible. For more information about agent nodes, see Adding agent nodes.
There is no single best pattern for agent collaboration. The right approach depends on the task, the agents’ abilities, and how much autonomy the system requires. As real-world use cases evolve, coordination strategies will continue to adapt and new patterns will emerge. Human expertise remains essential, as subject matter experts guide and shape how agents work together. Agent patterns become powerful when they are defined by using workflows, allowing human knowledge to drive coordinated AI behavior.
Agentic solution pattern using HR and healthcare examples
The following figure shows two examples of agentic solutions. In the HR example, the HR agent works with two collaborator agents that handle Travel and Pension requests. The Travel agent uses a mix of agentic workflows, tools, and other agents to process different user requests.
In the Healthcare example, the Healthcare agent uses an agentic workflow to classify the user request and then routes the request to the appropriate agent.
These examples show how agents and agentic workflows can work together to apply probabilistic and deterministic behavior where each is most effective. Both solutions use a supervisor agent pattern. In the HR scenario, the agent provides flexibility. In the Healthcare scenario, the agentic workflow provides predictability.
Create a set of linked activities and controls that are designed to achieve a specific business purpose or goal.
Create custom logic to insert into your flow using supported Python libraries, data types and data type conversions. See the various use cases for logic blocks.
Example videos on how to build an agentic workflow in watsonx Orchestrate.