Topic classification descriptions and routing
Key Takeaway
The agent picks a subagent by comparing the conversation with the names and classification descriptions of the subagents still available to it. When routing goes wrong, write detailed, unique classification descriptions that say which user requests should trigger each subagent.
An Agentforce agent can do several jobs. Each job is a subagent. Classification is how the agent picks the subagent that fits what the user wants. Routing is how the conversation gets to that subagent. Clear subagent names and classification descriptions drive most of that choice. When an LLM guess is not good enough, filters and deterministic transitions let you control routing.
Subagents
Beginning in April 2026, agent topics are called subagents. The functionality did not change, so topic classification and subagent classification are the same process under the old and new names.
A subagent's parts include a name, a classification description, reasoning instructions, and reasoning actions. By default, the name and classification description are used to decide when to use the subagent. The reasoning instructions and reasoning actions guide what the agent does after it moves into the subagent.