Sequential Chat
As a sequence of chats between two agents chained together by a mechanism called carryover, the sequential chat pattern is useful for complex tasks that can be broken down into interdependent sub-tasks.
The figure below illustrates how this pattern works.
In this pattern, a pair of agents start a two-agent chat, then the summary of the conversation becomes a carryover for the next two-agent chat. The next chat passes the carryover to the carryover parameter of the context to generate its initial message.
Carryover accumulates as the conversation moves forward, so each subsequent chat starts with all the carryovers from previous chats.
The figure above shows distinct recipient agents for all the chats, however, the recipient agents in the sequence are allowed to repeat.
To illustrate this pattern, let's create a sequential workflow to create our lesson plan for the fourth grade class.
A teacher agent will engage three agents sequentially:
- A curriculum designer will select a topic given the subject from the teacher
- A lesson planner will design the lesson with two iterations given feedback from the teacher
- A formatter will format the lesson
import Example from "/snippets/python-examples/sequentialchat.mdx";
The sequential chat is triggered by the teacher agent's initiate_chats method, which takes a list of dictionaries where each dictionary represents a chat between the teacher and the recipient agent.
The maximum number of turns in each chat can be controlled with the max_turns key. Each chat can also terminate before the max_turns, see the Ending a Chat topic for further information.
The result of the initiate_chats method returns a list of ChatResult objects, one for each chat in the sequence.
Different senders#
In the above example, the teacher agent was the sender in each chat, however you can use the high-level initiate_chats function to start a sequence of two-agent chats with different sender agents. See this notebook for an example.