For years I have had every figure about my money and not once a story I could argue with. So I built one on JarvisCore. You hand it a bank statement, the locked PDF exactly as it arrives, and five agents pick it up in turn: the Proposer reads the whole thing and suggests the categories that fit my life rather than a generic chart of accounts, the Categoriser places every transaction into them, the Watchdog says what departs from a place's own pattern, the Coach writes the plain account of where the money went, and the Auditor separates cost from waste. What made it hold together is that none of them know each other's names. The Coach does not call the Auditor, it asks for whatever offers 'avoidable_cost' in the mesh and waits, so I can swap an agent out and nothing else has to be told.Thanks to the inbuilt memoryOs once you correct one transaction placement and every payment to that place, past and future, follows it. The code is public, check it out for more info https://github.com/Prescott-Data...
What I like most is that there's no master agent to babysit. Agents find each other and talk directly, so adding another one is just starting it up. It's also a relief that credentials stay out of the agents entirely.
Happy to see this finally go live! 🙌
"never with their own keys" is the right instinct, moving auth to a broker instead of scattering secrets across every agent. but doesn't that broker become the one thing that's now a single point of failure/attack in a supposedly peer-to-peer mesh? if it goes down or gets compromised, do the agents just lose the ability to call out, or is there a fallback path that doesn't quietly reintroduce static keys?
@galdayan that's some good oversight and we definitely need to think about what happens when the broker is down and how agents react to that reality. Ideally because we want zero-trust I see us taking the direction of nuking and human in the loop as a response. But this is just me thinking on my feet. Thanks for pointing that out.
@askmuyukani fail closed and page a human makes sense for the failure mode. the part I'd worry about more is the compromise mode though - if someone gets into Nexus itself, fail-closed doesn't help you, it just means the attacker has the single most valuable credential store in the whole mesh instead of one agent's key. does the broker get the same "least privilege per task" treatment the agents do, or does it hold broad standing access to everything so it can hand out any credential on demand?
@galdayan yes same least privilege per task, so in production you maintain your credential vault could be secrets manager if deploying in google cloud or hashicorp vault (whichever it is) then enforce least privilege at that boundary, then interface this with the broker. Initially, we thought about having credential store within Nexus but when we thought about it we saw the risk you are talking about and secondly, it did not make sense "reinventing the wheel"
@galdayan aah I get your question now, yes yes it is ephemeral
Our CTO wrote a beautiful piece explaining the details of the e2e handshake here if can take a read https://medium.com/@mwenyinyo/the-nexus-protocol-standardizing-identity-and-connection-orchestration-for-autonomous-agents-5fa0fb7f36d9