From what I’ve read, the multiple AIs generally do have a result, just not a good one.

medium.com/…/two-ais-talk-to-each-other-the-resul…

tomsguide.com/…/what-is-gibberlink-why-its-freaki…

www.sciencedaily.com/releases/…/240318142438.htm

But perhaps someone has a link to positive outcomes.

what you’re proposing requires them to reason and understand each other. LLMs don’t do that, they take text input and construct an output based on words (tokens) that they have mapped to be close to the ones you entered into your prompt.

it’s a clever way to produce a plausible response, but it’s not thinking or reasoning.

A conversation/collaboration… not really.

You can create a ‘swarm’ of agents with differing roles, define different roles and phases, to have it iterate on a problem.

1. Investigations to discover and provide a condensed context of discoveries
2. Use this to propose a plan
3. Some number of iterations of one or more reviewing agents (you can give each an area of focus) criticising the plan, one or more agents to propose improvements based on the reviews, and an agent to review and apply the proposals before the next iteration

Groups of agents of the same role, operating in parallel, should ideally be using different models (or have context that gives them differing goals - eg focused on maintainable abstractions, security, scalability, test case identification, etc).

The implementation can do a similar thing - a code generator followed by reviewers, proposals for action, and then apply improvements… and you can iterate on testing or benchmarking too, all before hand-over.

This can improve results (at a non-trivial cost sometimes, so budgets are important) and it will still miss sometimes. You can help it of course with hints, directions or even implementations or stubs of implementations of abstractions you expect.