Workflows
Workflows orchestrate complex, multi-step processes by connecting agents, conditions, and transformations into executable graphs. When a single agent is not enough, workflows let you coordinate multiple agents, add decision points, and include human approvals.
Why Workflows?
Some processes are too complex for a single agent:
- Multi-stage approval -- Review, manager approval, finance sign-off, then execute
- Data pipelines -- Extract, transform, validate, load
- Customer journeys -- Qualify, enrich, route, assign, notify
- Conditional logic -- High-value leads take the premium path; others take the standard path
Workflows give you visual process design, conditional branching, parallel execution, human-in-the-loop approvals, and error handling with retries.
Sample Workflow
Here is how the 8 node types come together in a typical flow:
Core Concepts
Node Types
Nodes are the building blocks of a workflow:
| Node Type | Purpose |
|---|---|
| Start | Entry point (exactly one required) |
| Agent | Execute an agent with input/output mapping |
| Condition | Branch based on expressions |
| Parallel | Execute multiple branches concurrently |
| Merge | Join parallel branches back together |
| Transform | Transform data without an agent |
| Delay | Pause execution for a set duration |
| Human | Wait for human input or approval |
| Loop | Repeat a sequence until a condition is met |
| Webhook | Wait for or send an external HTTP callback |
| Coding | Execute code in a sandboxed environment |
| Notify | Send a notification (email, Slack, in-app) |
| End | Exit point (at least one required) |
Edges
Edges connect nodes and define execution flow. Every node (except End) must have at least one outgoing edge, and every node (except Start) must have at least one incoming edge.
Merge Strategies
When parallel branches converge at a Merge node, you choose a strategy:
| Strategy | Behavior |
|---|---|
| All | Wait for every branch to complete |
| Any | Continue when any single branch completes |
| Race | Continue when the first branch completes; cancel the others |
Input/Output Mapping
Data flows between nodes using JSONPath references. You can reference workflow input, previous node outputs, and nested fields to wire data through the graph.
Building Workflows
Visual Builder
The easiest way to create workflows is the Visual Orchestration Builder (Growth plan and above):
- Navigate to Orchestration > Builder in the sidebar
- Drag nodes from the palette onto the canvas
- Connect them by dragging edges between nodes
- Click each node to configure it in the right sidebar
- Click Validate to check for structural errors
- Click Deploy to make the workflow executable
The builder supports 8 node types: Trigger, Agent, Condition, Handoff, Human Approval, Parallel, Loop, and End. It follows BPMN-inspired visual conventions (circles for start/end, diamonds for decisions, rectangles for tasks).
Workspace and Project Scoping
The Visual Builder adapts to where you open it:
-
In the Admin layout, the builder is scoped to your currently selected workspace. Agents are grouped by team within that workspace, and flows are saved against the workspace. This is the right choice for orchestrating agents that all belong to the same workspace.
-
In the Projects layout, the builder is scoped to the project and can include agents from every workspace in the project. Agents are tagged by their origin workspace so you can tell them apart at a glance. When agents from different workspaces are connected, the platform automatically wires SLIM messaging between them on deploy so they can communicate across workspace boundaries.
Cross-organization flows (agents from federated external tenants) are also supported in project-scoped flows through dedicated federation nodes. These require an existing SLIM federation trust to be in place.
Four starter templates are available: Sales Pipeline, Support Triage, Content Pipeline, and Code Review. Start from a template and customize rather than building from scratch.
Workflow Settings
| Setting | Description | Default |
|---|---|---|
| Max Execution Time | Maximum total workflow duration | 1 hour |
| Max Node Retries | How many times to retry a failed node | 0 |
| Enable Logging | Detailed execution logs | On |
| Share Memory | Agents share memory context across the workflow | Off |
Validation Rules
Workflows are validated before they can be saved or deployed:
- Exactly one Start node is required
- At least one End node is required
- No orphaned nodes (all must be connected)
- All edges must reference valid source/target nodes
- No circular dependencies (except explicit loops)
- Cycle detection prevents infinite loops
Running Workflows
You can execute a workflow manually from the dashboard, via a trigger, or from another workflow. Each execution creates a run with its own status tracking.
From the dashboard, you can monitor:
- Current node and execution path
- Completed and pending nodes
- Output from each node
- Errors and retry attempts
You can also cancel a running workflow at any time.
Import and Export
Workflows support multiple import/export formats:
| Format | Direction | Description |
|---|---|---|
| BPMN | Import | Import from standard BPMN tools |
| Agent Spec | Import/Export | Oracle Agent Spec format |
| Team Manifest | Import | Create visual flows from team manifests |
| YAML | Export | Export as team manifest YAML |
Best Practices
Begin with a linear workflow (Start, Agent, End) and add complexity as needed. Validate at each step.
Rather than building separate workflows for different scenarios, use Condition nodes to branch within a single workflow.
Any step that involves money, customer data, or irreversible actions should include a human approval gate. See Human-in-the-Loop.
Always configure a max execution time. Without one, a stuck workflow can consume resources indefinitely.
Next: Learn about Human-in-the-Loop for human approval workflows.