Agent action descriptions and instructions
Key Takeaway
An agent's LLM reads action names and instructions to choose which action to run and what data to pass, so write them clearly and with relevant detail. Keep sensitive or deterministic business rules out of the instructions, and build them into the reference action instead.
Agent action instructions are plain-language descriptions that tell an agent what an action does and when to use it. Each input and output of an action also has its own instructions. The agent's large language model (LLM) reads these names and instructions to choose which action to run and what data to pass to it. Clear instructions make an agent behave more predictably.
How an agent chooses an action
An agent picks actions by reasoning, not by fixed rules. It uses an LLM and the reasoning engine to search all the subagents and actions assigned to it. Then it selects one or more actions to run, based on the user's request.
The names and instructions of each action, and of its inputs and outputs, decide when and how the agent uses that action. The agent uses both the action instructions and the subagent instructions to select an action.
Keep no more than 10 actions in a subagent and no more than 10 subagents on an agent. With too many, the agent can choose the wrong one.