Using Nimble Web Search Agents: pick research, enrichment, or dataset-building as the use_case before creating the run, because it's locked after creation and can't be changed mid-run. Create the run, then either subscribe to live progress events or poll status until result is ready. Only reach for an agent when the source is unknown, data is scattered across inconsistent pages, or you need free-text synthesis; for a fixed known URL use an Extraction Template, and for raw results use plain search. Track the run ID and poll with backoff; don't fire duplicate runs when one is already in flight.

Context: Web (Nimble agent-skills reference for Web Search Agents): documents POST /v2/agents/* for open-ended web research where the source isn't known upfront. Gotchas: use_case (research, enrichment, dataset building) is set once per agent run and then locked; the run lifecycle is create -> status -> result; you can subscribe to live events or poll as a fallback. Requires Nimble CLI 1.2.0+. When a single known page can be parsed directly, prefer an Extraction Template over an agent run.