- Understanding Workflows
- Why Workflows Matter
- For example: Flight Data Processing Pipeline
- Create your first workflow
- Next Steps
Workflows in Solace Agent Mesh provide a powerful way to orchestrate multiple agents through explicit prescriptive workflows rather than relying on dynamic LLM driven workflow generation. Unlike the orchestrator agent which uses LLM reasoning to dynamically determine how to accomplish tasks, workflows follow predefined execution paths that you specify in configuration. Each step, branch, and iteration is explicitly defined, giving you complete control over the agent interaction sequence.
Workflows are particularly valuable when you need predictable execution that follows the same path every time, business processes with compliance or audit requirements, visual representation of agent interactions in the UI, and fine-grained control over error handling and retries. This makes workflows ideal for scenarios where consistency, auditability, and deterministic behavior are critical requirements.
Workflows bring several important advantages to your agent mesh architecture. They provide transparency by explicitly documenting the sequence of operations in a human-readable format, making it easy to understand exactly what will happen when a workflow executes. This transparency is crucial for compliance and audit scenarios where you need to demonstrate that specific procedures were followed.
Workflows also enable reliability through built-in retry logic, timeout controls, and error handling at both the workflow and individual node levels. You can define exactly how the system should respond when things go wrong, ensuring graceful degradation rather than unexpected failures. Additionally, workflows register themselves as discoverable agents, allowing other agents and the orchestrator to invoke them just like any other agent in your mesh.
To learn more about workflows, check out the Solace Agent Mesh Workflows official documentation
Consider a real-world aviation scenario where you need to process flight data through multiple validation and enrichment steps. A workflow can orchestrate the entire pipeline:
- Validate Flight Data: An agent verifies that incoming flight records contain all required fields and follow proper formatting.
- Enrich with Weather Data: Another agent queries weather conditions along the flight route and adds this context.
- Check Compliance: A specialized agent validates that the flight plan adheres to FAA regulations and airspace restrictions.
- Generate Report: Finally, an agent produces a comprehensive report summarizing the validation results and recommendations.
This sequence must happen in order, with each step depending on the success of the previous one. If validation fails, there is no point in enriching the data. If compliance checks fail, the report must reflect this status. A workflow ensures this exact sequence executes reliably every time.
Workflows are defined using YAML configuration files with several key sections. Here is a simplified example based on the structure used in Solace Agent Mesh:
apps:
- name: flight_processing_workflow
app_module: solace_agent_mesh.workflow.app
broker:
# Broker connection details
app_config:
namespace: ${NAMESPACE}
name: "FlightProcessingWorkflow"
# Workflow timeout settings
max_workflow_execution_time_seconds: 300
default_node_timeout_seconds: 60
workflow:
version: "1.0.0"
description: "Validates and enriches flight data"
# Define expected input structure
input_schema:
type: object
properties:
flight_id:
type: string
description: "Unique flight identifier"
required: [flight_id]
# Define output structure
output_schema:
type: object
properties:
status:
type: string
report:
type: string
required: [status, report]
# Define the workflow steps
nodes:
- id: validate_flight
type: agent
agent_name: "FlightValidator"
input:
flight_id: "{{workflow.input.flight_id}}"
- id: enrich_data
type: agent
agent_name: "DataEnricher"
depends_on: [validate_flight]
input:
flight_data: "{{validate_flight.output}}"
- id: generate_report
type: agent
agent_name: "ReportGenerator"
depends_on: [enrich_data]
input:
enriched_data: "{{enrich_data.output}}"
# Map final outputs from node results
output_mapping:
status: "{{generate_report.output.status}}"
report: "{{generate_report.output.report}}"-
Workflow Metadata: The
versionanddescriptionfields document what the workflow does and help with versioning as your workflows evolve over time. -
Input and Output Schemas: These define the contract for your workflow using JSON Schema. The
input_schemaspecifies what data the workflow expects to receive, whileoutput_schemadefines what it will return. This validation ensures type safety and clear interfaces. -
Nodes Array: This is the heart of the workflow, defining each step in the process. Each node has:
id: A unique identifier used to reference this step's outputtype: The node type (agent, switch, map, loop, or workflow)agent_name: Which agent to invoke for this stepdepends_on: An optional array specifying which nodes must complete before this one executesinput: The data to pass to the agent, using template expressions like{{workflow.input.field}}or{{previous_node.output.field}}
-
Template Expressions: Workflows use
{{...}}syntax to reference data.{{workflow.input.flight_id}}accesses the initial input, while{{validate_flight.output}}retrieves the output from a completed node. This creates a data flow through your workflow. -
Output Mapping: The final section maps node outputs to the workflow's overall output. This allows you to combine results from multiple nodes or extract specific fields from the final step.
-
Timeout Settings: The
max_workflow_execution_time_secondsprevents workflows from running indefinitely, whiledefault_node_timeout_secondssets a timeout for each individual step. These can be overridden per-node if needed.