Prompt and instruction maintenance in Agent Builder
Key Takeaway
Make small changes to an agent's instructions and retest after each one. Put sensitive or deterministic rules into the action itself, because instructions don't replace coded business rules.
An Agentforce agent follows written instructions and sends system messages. Admins maintain these in Agentforce Builder: they edit a draft version, test it in the Preview panel, and then commit and activate it. A large language model (LLM) interprets the instructions, so maintenance is a loop of small changes followed by retesting.
Layers of instructions
Instructions come in layers:
- Agent-level instructions guide how the agent interacts in all situations, including rules it should follow.
- Subagent instructions give the agent context for one subagent's job.
- Action instructions, including input and output instructions, describe what an action does and what data to use.