Final response validation for Service agents
Key Takeaway
An Agentforce Service agent validates its proposed response before it sends it. If the response fails, the agent replaces it with a new one, which can happen after the response has started streaming to the user.
Final response validation is the last check an Agentforce Service agent runs on its proposed response before it sends that response to the user. It is an extra step between reasoning and sending, at the end of the path that every user message follows through the Atlas reasoning engine.
The path a user message follows through the Atlas reasoning engine
Every user message follows a set path through the Atlas reasoning engine:
- The agent goes to the starting subagent, which is the subagent marked with the
start_agentprefix in Agent Script. By default, this is the Agent Router. The Agent Router picks a subagent based on recent conversation history and the subagents available to the agent. - The agent resolves that subagent's reasoning instructions top to bottom, in the order they're written. No LLM is involved yet. If an instruction transitions to another subagent, the agent redirects the conversation before any reasoning happens.
- The agent builds a prompt that holds:
- the resolved subagent instructions
- agent-level instructions
- recent conversation history
- the reasoning actions the subagent allows
- The agent reasons with the LLM. The LLM decides whether to:
- run an action
- ask for more information
- ask a clarifying question
- respond directly
- After the agent runs an action, it checks the result in a reasoning loop. It then decides whether to run another action, ask more questions, or send a final response. The agent can go through up to seven reasoning loops.
- Final response validation checks the proposed response. This extra step applies only to Agentforce Service agents. Other agent types, such as Agentforce Employee agent, don't have this step in their path.
- The agent sends the response to the user.