Multi Agent AI Personas

Tom: Explains intro, but I missed it

PJ: Talked about having a UI that you can talk to different agents.

Max: Specialized agents trained on different data could talk to each other.

Anton: How do you create a persona

Tom: You tell the agent to behave like a persona to don't actually train it

Xing: It's about refining a base model rather than training a new model

Max: AI's generally agree with you. In order to get them to critique each other we'd need to prompt them to be critical of each other.

Tom: After an long painful interaction I've tried asking what was the prompt I should have used to have got there qiucker. However the AI often just tells me that I already took the quickest route. Tom: We should be experimenting now because my CEO thinks there will be a step change in pricing.

Tom: We tried using a single AI with a prompt telling them to assume two personas but it wasn't succesful. They didn't switch persona cleanly.

Tim: Crew AI - already does something similar

Ade: We have a lot of customer data we could use to make specialist AI's but we don't have legal persmission to use it.

Everyone: GDPR concerns. User data is being sent to LLM's

Graham: Can you include a human in the loop to evaluate responses and add some reinforcement learning

Tom: Gell Mann Amnesia Effect. When you are an expert it produces low quality content.

Anton: How can you validate an agent?

Tom/Max: You have a persona that forces agents to validate their sources. This still does not eliminate false positives but reduces the probability.

Anton: If you still have to check I don't see the value for anything that it is critical to get right

PJ: The utility is that everyone can ask the CEO questions at the same time.

Tom: If that frees up 10x the CEO's time and he can validate the responses in less than that, it is still net positive for the CEO's time.

PJ: We replaced functional tests with metrics and rolled back deployments quickly when there was a problem. Perhaps you can do the same with unverified decisions.

Graham: You'd have to apply this not only revenue metrics but also security and other dimensions that might be hard to measure.

Tom: Can agents maintain agent generated code

PJ: When I'm vibe coding I often ask the agent to critique and refactor the code. This could be done by another agent.

Graham: If HR is represented by an agent this could be legally difficult.

Tom: The agent is only an advisor the human has to be accountable

Xing: When humans are in a room and have to come to consensus they usually make good decisions. Would the same not happen with agents?