The Everywhere Computer

I'm looking into the dynamic of giving AI agents conditional access to data. Couple of areas that I'd like input:
- users can grant an agent access to their own private data
- agents can share private data to each other
- agents can share vectors/weights/learning with each other

Thoughts? Experiences? Resources?

I like Modular Synthesiser, MiniGrids, and Go (Baduk).

Also, I'm a developer at https://threshold.network/

https://github.com/theref

$THRF

Driving growth with passion | /wevibe  | built @degenbott

let's create an ai agent groupchat and let them ramble on freely.

Co-founder, Scout Game. 
https://scoutgame.xyz

First option :) Maybe my personal health tracker agent can interact with yours when we want to do a joint workout. Maybe companies want their agents to interact as part of contract negotiations

Does that second bullet mean that my personal agent and your personal agent can share private data? 

Or that my agent 1 can share it with my agent 2?

Designing and building AI systems

https://agents.gladio.ai

I guess I saw weights sharing as some sort of "learning/training sharing" 

And yes, you've absolutely nailed it. I just wondered how people are going about that at the moment, with a mind to future inoperability. How do I encrypt my data and define access conditions for agents that don't even exist yet

Given my understanding, I don't see how / why weights sharing would be useful or feasible, given it's super heavy and forms a core part of the models' ability to interact with the world.

But segregated agent swarms where each agent has access to a subset of your data world?

That makes sense and is the way forward in 2025.

Here's the technique you're looking for; give and revoke data access to agents via tools, manage that process outside the agent itself.

You should never depend on agents for enforcing your security system. The agent is a consumer of RABAC permissions just like any human user.

Achieve inter agent data sharing by agent <> agent messaging