Learning without a teacher

[rickyloynd-microsoft]

Agents will be more easily taught by users if they can also learn from their own experience.

Human students don’t need their teachers to guide them through every detail of a task, since they can quickly learn many things on their own. Similarly, a teachable agent should not require users to spell out all the edge cases that a skill needs to handle. An agent’s ability to quickly learn from its own experience will be especially useful in group-chat settings where the user may not even be in direct communication with a teachable agent. And this ability will help the agent learn from user instructions that are vague, partial, or distributed over multiple dialogue turns.

• Coding scenario to be addressed AutoGen’s AssistantAgent can currently correct errors that appear when the UserProxyAgent runs code. But AssistantAgent never learns to avoid making those very same errors in future chats. Users shouldn’t need to tell the agent not to repeat mistakes that caused failures. An agent’s inability to learn such basic things on its own frustrates users and increases LLM charges.

[thinkall]

We are closing this issue due to inactivity; please reopen if the problem persists.