Agent-level instructions versus topic instructions
Key Takeaway
Agent-level instructions are in the prompt the agent sends to the LLM no matter which subagent is handling the conversation. A subagent's instructions give rules and context only for the job that one subagent handles.
Agentforce agents get written guidance at two levels. Agent-level instructions apply to every conversation and set the rules the agent always follows. Subagent instructions give rules and business context for the job one subagent handles. Both levels end up in the prompt the agent sends to the large language model (LLM). As an administrator, you edit agent-level instructions in Agentforce Builder, and you test instruction changes by chatting with the agent in Preview.
Agent-level instructions and subagent instructions
Beginning in April 2026, agent topics are called subagents, with no change in functionality. So topic instructions now appear as subagent instructions. You may see both terms during the transition.
The two levels differ in what they tell the agent and what they apply to.
| What they tell the agent | What they apply to | |
|---|---|---|
| Agent-level instructions | How to interact in all situations, including any rules it should follow | Every agent interaction, no matter which subagent is handling the conversation |
| Subagent instructions | Guidelines for agent behavior, and the context the agent needs to do one job | Only the job that one subagent handles |
Examples of subagent instructions include:
- Definitions of business terms
- Decision criteria for choosing between similar options
- Which action to use in which situation, for example "Use the Track Order action when customers ask about package location or delivery dates."