Hold a conversation

A conversational agent keeps one run open while a person sends several messages. Each reply goes to the same run, so the agent remembers what was said.

Set it up

1. On the agent, set Wait for Reply (Hours) to how long the run waits for the next message, from 1 to 168.
2. Build a screen Flow for the record page:
   1. On the first screen, ask for the person’s message.
   2. Start the run with Start Agent Run. Pass the message as Input and {!recordId} as Anchor Record ID.
   3. Read the answer with Get Agent Run Result. While Status is Running, show a screen with a Refresh button that loops back to this step. The answer is ready when Status is Waiting.
   4. Show the answer. Ask for the next message.
   5. Send it with Reply to Agent Run, then read the answer again.
3. Add the Flow to the record page in Lightning App Builder.

Rules

• Only the person who started the run, or an operator, may reply.
• The limits on the agent, such as Max Model Calls, apply to each reply on its own.
• The run ends with its last answer when the wait runs out. Start a new run to continue after that.
• Transcript Turn Limit decides how many earlier messages the model sees. Older ones are dropped first.
• Reply to Agent Run works only while the run waits for the person. While the agent is still answering, it returns “This run is not waiting for a person’s reply.” Read the answer first.

Cost

Each reply is at least one model call. On an agent with an identity, each reply also costs 2 jobs and 1 API call to sign in as the execution user.