Agents are great at open-ended work. When a process has to run the same way every time, I reach for a workflow: a graph of typed nodes that Solace Agent Mesh runs in dependency order. There are several node types in workflows.
| Node | What it does |
|---|---|
agent |
Sends a structured request to an agent and captures the response |
tool |
Calls one tool directly, no LLM involved |
switch |
Picks one downstream node based on a condition |
map |
Fans a node out over a list of items |
loop |
Repeats a node until a condition is false |
workflow |
Runs another workflow as a step |
decision |
Asks a decision model typed questions and routes on the answer (Early Access) |
Parallelism comes from depends_on: nodes that don’t depend on each other run at the same time, and a node that depends on several waits for all of them.
Here is a shipment exception triage workflow. One agent classifies the exception, a switch routes it to one of four recovery agents, two other agents run in parallel, and everything joins back up for the report:
And here is a peak season plan using map (one planning step per warehouse) and loop (refine until coverage hits 95%, at most three times):
The YAML for those two nodes is short:
- id: plan_sites
type: map
depends_on: [forecast_volume]
items: "{{forecast_volume.output.sites}}"
node: plan_site
- id: refine_network_plan
type: loop
depends_on: [plan_sites, market_scan]
node: review_plan
condition: "{{review_plan.output.coverage_pct}} < 95"
max_iterations: 3
Always set max_iterations on a loop. The default is 100, and that is a lot of LLM calls if your condition never flips.
When I forget a field, sam config schema show workflow --type loop (or any other type) prints exactly what that node accepts.
Docs:
Workflows are Early Access, so this is a great time to shape them. Download the Solace Agent Mesh desktop app and build one.
What process would you model as a workflow first? Share it below and I’ll suggest which node types fit.