10 Comments
User's avatar
Marcus Chang's avatar

Giving each agent its own embedded engine instead of routing through one shared warehouse is a clever inversion of the usual approach.

I've been mapping out agent architecture patterns like this in a repo of runnable examples, worth a look.

https://github.com/hardness1020/awesome-agent-architecture

Jim Ryan's avatar

That is mind-blowing. Cool stuff

Steven Yampolsky's avatar

I think you updated something in the article and all the code blocks became `code…` vs the multiline blocks you had that didn’t leave things open to interpretation…that or the settings on my substack app changed

Steven Yampolsky's avatar

And it’s back! Maybe was the app?

Joe Reis's avatar

No idea. Substack glitches often

Dragos Crintea's avatar

I think this is one way, the other way I was exploring this was from a different angle. I was thinking that I do not need an entire RDF or OWL to represent that and I built in semantido a concept registry. It's a purely vibe-coded experiment after burning $50 with Fable but based on a hypothetical experiment implemented on top of what semantido already supported.

You pinned out yourself that a hash does not prove that every relevant semantic difference has been captured, and formatting changes produce false hash differences that the normalizer has to paper over.

What a context registry does is declarative and positive: Counterparty_EMIR DISTINCT_FROM Counterparty_MiFIR is an assertion of intent, made once by someone with domain knowledge, that survives independently of any particular physical artifact or hash. You discover the disagreement empirically every time two contracts collide, semantido encodes it as knowledge upfront.

Here are both references, the concept-registry and a simulated conversation between agents.

https://semantido.ai/guides/concept-registry/

https://github.com/hikarilabs/semantido/tree/main/examples/04_federated_agents

Sadie St Lawrence's avatar

This is really helpful!

Joe Reis's avatar

Thank you, thank you