Action instructions and input instructions for action selection
Key Takeaway
An agent picks a subagent first, then picks actions inside that subagent by reading names and instructions. When it runs the wrong action, find out which of those two steps went wrong, then fix the wording for that step.
Action instructions are short, plain-English descriptions that tell an Agentforce agent what an action does and when to use it. Input instructions tell the agent what data an action needs and how to find or format that data. The agent's large language model (LLM) reads action names and these instructions to decide which action to run and what to pass into it, so clearer wording gives more predictable behavior.
How an agent chooses what to do
An agent chooses what to do in two steps. First, it matches the user's request to a subagent by comparing the last six turns of the conversation with each subagent's name and classification description. Then, inside that subagent, it uses the subagent instructions plus the names and instructions of the subagent's actions to decide which actions to run.
Before April 2026, subagents were called topics. The rename did not change how they work.
In the second step, the agent uses the LLM and the reasoning engine to pick one or more actions to run. The names and instructions of an action, its inputs and its outputs together decide when and how the action is used.
Salesforce recommends no more than 10 actions per subagent and no more than 10 subagents per agent. With too many, the agent can choose the wrong one.