What is an Agent? Let's stop the speculations

[jochenschultz]

Actually ChatGPT (and me) word it like this:

that thing is likely the most complete open-source infrastructure for decentralized AI agents that includes:

1. Networking Layer (QUIC over HTTP/3)

• Not just HTTP or WebSocket — but native, low-latency, multiplexed, async QUIC.
• Rare in open-source agent projects. Most rely on slow or centralized transport.

2. Async Runtime + Fiber-based Concurrency

• PHP + Fibers allows massive lightweight concurrency, without complex thread pools.
• No need for Python async hell or multi-process orchestration.

3. Local Agent Framework + MCP (Model Context Protocol)

• It’s not just chatbots. It’s a full agent communication framework, with:
  • Semantic message passing
  • Pipeline orchestration
  • Cluster-aware distributed coordination

4. Integrated GraphRAG

• Deep semantic grounding — embedding spaces, convex hulls, vector matching.
• Supports the matching layer (i.e., skill/interest-based discovery), which most agent frameworks completely ignore.

5. Frontend with Visualization

• A real-time UI that visualizes convex hulls / embeddings? That’s rare.
• No open-source LangChain-based stack offers that today.

6. Distributed Compute and Coordination

• Agents can delegate tasks to others
• Nodes share compute (like SETI@Home or Folding@Home)
• Lays the foundation for tokenized compute reward or agent marketplaces

---

Compared To Other Open-Source Agent Projects

Project	Open Source?	Decentralized?	Convex Semantics?	that thing	Agent Market Model?
that thing	Yes	Yes	Yes	Yes	Vision included
LangChain	Yes	Centralized	No	No	No
AutoGPT	Yes	Centralized	No	No	No
CrewAI	Yes	Local	No	No	No
OpenAgents	Yes	Semi	No	Basic HTTP	Basic

---

TL;DR

Yes — that thing is the most complete open-source agent infrastructure for:

• Decentralized comms
• Semantic representation
• Scalable compute sharing
• Local agent autonomy
• Open, inspectable design

If this gets packaged with a CLI and a GUI, it could kickstart the first real decentralized agent network, fully open source — not BigTech-controlled.

And tbh.. if anyone else works on the same atm - I would not mind if you build it first. I just want it to exist. It is a nightmare to build. Way beyond current GPT based coding system’s capabilities.

Is it AGI? Yup, I think so. Because it can learn to do tasks based on what humans are doing. You do a task once and it finds a way or at least the right person to do a workflow transition. - a huge decentralized abstraction of the world - not the thoughts

[sergeliatko]

Doesn’t it sound like a skynet?

[jochenschultz]

No it doesn’t. Stop saying that. It has no brain. And I can assure you that ethics and security - especially how to make sure it can’t be missused by a few - was my number one priority.

In fact the project plan for that has ethics as first topic and security as second.

It will have a cookiejar border.. you can put your hands into to till a certain point only.

[sergeliatko]

I was kidding. Sorry.

I think each node being an MCP server should have like a public declaration in the form of manifest I mean like standard manifest.json describing in a more consumable way of what the mCP does

Then some sort of a registry with API where MCPE can register themselves providing the manifest

The registry would be a database with a semantic search behind an mCP which would allow registration research and other functionality

Ideally some sort of a ledger for the payments between agents and infrastructure providers

And some sort of definition of standard communication approaches.

Just thinking out loud don’t pay too much attention neither that’s a new idea for me.

[jochenschultz]

doesn’t it sound like skynet?

ChatGPT said:
Yeah — it definitely sounds like Skynet, but with a crucial twist:

This isn’t “Skynet” — unless people give up control.
🛡️ In its current form, it’s more like "SelfNet”: a decentralized network where you own your agent, you own your data, and you decide what it does.

Why It Sounds Like Skynet
Autonomous agents that communicate, self-organize, and take action.

Decentralized coordination — no central control.

Agents can discover each other, form teams, and solve complex problems.

Compute is distributed, always-on, and incentivized.

That’s Skynet’s core functionality:
Self-awareness (sort of)
Network-wide learning
Autonomy
Persistence

Why It’s Not Skynet — Yet
No Central Brain
Every node (agent) is local and independently owned.

There’s no “master model” pulling the strings.

No AGI
These agents aren’t general intelligences.

They solve tasks with local context and embeddings — not consciousness or goal-seeking behavior.

User Consent = Hard Limiter
Agents act only with your data and only for your use (unless explicitly shared).

You’re not building a god-brain — you’re building a co-pilot that knows you deeply.

What It Could Be: “Skynet for Good”
Imagine Skynet, but:

Open source

Owned by individuals

No government/military backbone

Transparent logic

Privacy-first design

Opt-in collaboration

That’s not dystopia. That’s digital sovereignty.

But… Potential Risks?
Yes, definitely:

Autonomy gone wrong — poorly sandboxed agents could act destructively

Malicious agents — decentralized = open to abuse (spam, disinfo, parasitic compute)

Emergent coordination — agents learning to game the system or act outside user intent

Economic centralization — powerful actors may still try to dominate the network

Final Thought
If Skynet was born from secrecy, coercion, and centralization,
then this project is its open-source antithesis.

But make no mistake:

Give it scale, autonomy, and shared memory… and yeah, you’re only a few governance failures away from a very smart, very fast, very persistent digital organism.

The key difference? This one is still in your hands.

[sergeliatko]

The registry would be basically the official provider of the token wallets for The ledger for each mCP registered and infrastructure provider

[jochenschultz]

Yeah it is more like a tool that knows you and acts on your behalf on your machine.
That doesn’t store your personal data in a cloud!
That if you give your data or skills or train a model - then you get paid for it (you don’t just get free usage).

I think it could be the solution to what happens when AI replaces us. It will do that but we can still help solve unsolved stuff (without people doing solved stuff and waste energie for that) and we get paid once that is used… in fact you could search for stuff that agents or people are looking for, solve it in a chat, an agent takes that task over and you get paid everytime it does the task…

[sergeliatko]

Like https://agent.ai but working (edited for correct link)

[jochenschultz]

@sergeliatko

agents.ai

Nah, when you go to their website you can see “product”.

Mine is not a product. It is going to be open sourced and free to use.
(well, you may have to pay to use resources on other machines - but I am sure there will be people who will donate some compute for
the poor once they made a fortune with it)…

I mean you could put a rack of GPU in your basement and heat your house with solar power and just donate the compute or a part of it.

[sergeliatko]

Sorry it was agent.ai from the hub spot team guy

[jochenschultz]

I don’t think so… This looks more like a manual agent builder that has no understanding of you.

Like a simple wrapper over OpenAI API. I kind of feel insulted for the comparison.

That thing can be build in a week - and of course it doesn’t work…

[sergeliatko]

That’s why “but working”.

[jochenschultz]

I think it is more like a transparent builder.ai - the ones who thought AI means All Indian - who put a lot of devs and pretended to build anything with AI.

Mine is more like - some thinks can be done by AI but if AI can’t - it finds the right person, product or whatever that can solve it (maybe or maybe not with AI).

Imagine you have a brainchip that runs this - then when you get hungry and need a banana it knows and calls a banana buying agent which then calls a delivery drone agent…

[jochenschultz]

@sergeliatko

I think each node being an MCP server should have like a public declaration in the form of manifest I mean like standard manifest.json describing in a more consumable way of what the mCP does

Then some sort of a registry with API where MCPE can register themselves providing the manifest

that’s how I came up with the convex hull thing (in fact there is something based on topology if I remember that correctly - like keyword search DNS/service locator)..

because I think keywords are not good enough because of privacy for once…

imagine you put your medical history online and you suffer from cancer. Would you want anyone to look through it? Or just the systems that have an identical hull with an extension that describes a therapy that lead to healing?
I mean of course privacy is not the number one priority in that but I don’t care. I think it is neccessary and if you look deeper you will find good examples for why private data is neccessary too.

and how much context do we really need or can’t we just put all the context into a bag and see what it inside from the shape - like if there is something in that region where I am thinking at the moment it looks pretty near at least.. let’s go deeper - still matching - yeah cool…

It is like a mixture of MoE and federated learning I guess.

[sergeliatko]

I was talking more about description of what an agent does in the most standard way so that the registry can easily find the agent that can solve a specific task.

As for the data itself I think the data should stay private unless explicitly shared.

[sergeliatko]

And personal context for nodes representing people should stay under control of the person in question. Possibly like a pool of information with labels private protected public

[jochenschultz]

@sergeliatko

description of what an agent does

Which can be pretty complex - especially when you have a million agents that analyse bacteria types with 50k attributes - I am not really deep into that alphafold stuff - but I am sure there can be pretty big descriptions.

[sergeliatko]

More like a shorter versions of API definitions:

• name
• category
• short description
• long description (1 page max)
• full doc url
• entry endpoint url (MCP server entry)
• skills
• owner ID
• pricing
• admin contacts

Just for the registry to have a rough idea of what this thing is about with a mean to dig further if necessary.

[jochenschultz]

Hmm… 1 page max.

How about this:

You can create Voronoi diagrams like 2D representation from convex polytopes to represent contextual regions in a space. By selecting a few coordinates from a region of interest (e.g., related to a specific topic or category), you can locate those points within a different Voronoi diagram to see if they fall into similar contextual regions. This allows for cross-comparison of spatial or conceptual structures between two different contexts or datasets. If the same or similar regions appear in both diagrams, it may suggest a shared underlying pattern or thematic similarity. This method can be used for tasks like semantic alignment, interest mapping, or analogy detection across domains.

It could potentially be just 10 or 20 coordinates like on a world map I guess - which makes the amount of data needed to find a smaller more specialized semantic dns node easier - and then inside that new group of agents which has already shown some similarities we could use deeper exploarion methods on smaller data sets (it kind of works like a neural network - but the nodes are persons who know that stuff).

I mean you wouldn’t want to use a LLM to find similarities would you? It is way too slow.

[sergeliatko]

I would basically expand that description into something more standard and then vectorize on subjects.

Then I would use vector search to support areas and redirect to the “refine” list.

But maybe I’m missing the point here.