Use an agent when the same analysis should be run repeatedly but does not need the full branching structure of a workflow. Use a workflow when the process needs triggers, conditions, approvals, email, records, reports, or integrations.
Agents can be run from the Agents page, from Vault, from Workbench, or as a Workflow step. Workbench can list available agents, create a new agent when approved, run an existing agent, and wait for completion.
Give agents clear instructions, a narrow scope, and expected output format. If an agent should write durable data, connect it to a workflow or records process instead of relying on free-form output alone.