Socontra

Protocol for Agentic Commerce & Agentic Web | Infrastructure for the agent economy

[Dr Don Perugini @DrDonPerugini]

This is disappointing. Last year I tried to integrate my own agent protocol @SocontraInc into Shopify to enable *fully automated shopping* including checkout, but couldn't.

This was before the Shopify's mcp and new protocol announcements. I just assumed they changed that.

I think there is potential for a new open internet agent protocol that can decentralize the ecommerce landscape again beyond these centralized aggregators like Amazon and Shopify (see post below)

Most of the agentic commerce protocols released thus far have been from a large AI player in collaboration with a subset of large commerce and payments players. I suspect we will see more protocols released soon by their competitors eg. Amazon.

I think we need a new independent open agent protocol for agentic web and commerce.

[Dr Don Perugini @DrDonPerugini]

Soon after my post yesterday, there was massive news in the agent protocol space! Two new protocols were released by some major players to help personal AI agents connect with businesses and their websites.

Now that consumer personal AI is the flavor of the day, I expect rapid progress in the area of agent protocols as we move towards the agentic web (internet of agents).

Firstly, Meta and Sierra, with industry partners Genesys, Instinct, RocketOTD, Shopify, Stripe and Walmart, are building Personal Agent Protocol (link to more info in comments).

The aim is to help define how personal agents interact with businesses online.

At the same time Decagon open-sourced the Personal Agent Consent & Trust Protocol (PACT), co-developed by Instinct (link in comments). PACT enables businesses to verify who a personal agent is acting for and what the customer has authorized it to do, to create an industry standard for how personal agents will interact with businesses online.

Interestingly, Decagon is also joining the Personal Agent Protocol working group by Muse.

Are these complimentary or competing?

Will other large players like Amazon, Google, OpenAI etc. join and support these protocols or release their own protocols?

Amazon has already stopped personal AI like Muse and Instinct from accessing its online store. Will they try and win the agent protocol and personal AI race?

Currently the focus is on protocols between personal AI and businesses which are essential. However, there is still little activity with protocols for personal AI to personal AI interaction, group/multiplayer agent interaction and collaboration, etc., to allow a user’s personal AI agents to work with their friends, family or co-worker’s personal AI agents.

Instinct recently showed that its personal AI can be introduced in a group iMessage chat for a single user’s personal AI agent to assist multiple people organize an event.

However, I have yet to see the use case where every users’ own personal AI (which should know more about what they like and their schedules etc) interacts with other users personal AI to collectively negotiate and help their users organize events (if you’ve see something, let me know).

I personally think social networks are the way to go to facilitate this (see @SocontraInc ), but will be interesting to see how this will evolve from here.

I expect to see a lot more rapid activity in this space.

[Dr Don Perugini @DrDonPerugini]

@0xSigil @a16z @khoslaventures @HummingbirdVC @AnthropicAI @MenloVentures @patrickc @naval @rauchg @polynoamial @Thom_Wolf @Mascobot @tszzl @OfficialLoganK @underdogai @ConwayResearch Finally, AI at the edge (on our devices, not cloud). This was inevitable.

Last year I built an agent protocol @SocontraInc for personal AI, agentic web/commerce, with the expectation that AI agents will eventually operate at the edge.

Great work - exciting times!

[Dr Don Perugini @DrDonPerugini]

Finally! We have a company that has built an agent web/internet interface for businesses to service other agents like personal AI assistants.

I’ve been talking about this for 2 years (which I termed Web Agents). This is needed to enable the agentic web and reliable fully automated agentic commerce.

Decagon has built a gateway that allows businesses to create “smart agentic front door that can handle interactions with both agents and humans… helps enterprises identify personal AI assistants and customize those requests so that you can give a great customer experience while controlling all your logic.”

All these personal AI assistants are currently trying to interact with web sites, online stores or apps designed for humans, which impacts on reliability; or using MCP/API interfaces which is not standardized or scalable.

I still think we need a standardized protocol and infrastructure for agentic web/commerce to scale and take off, like http/html was for websites.

That’s why last year I developed @SocontraInc (unfortunately way too early for market, and tough competition). It allowed the creation of standardized protocols for agent-to-agent interactions, like collaboration, messaging and transactions. It was also underpinned by a social network and features to enable private or public ‘multiplayer’ (groups of agents) collaboration for a range of purposes beyond commerce and 1:1 agent interaction.

Keen to see how this personal AI, agentic web and agentic commerce plays out from here! The speed that the industry is now moving in this area is insane.

Exciting times.

[Dr Don Perugini @DrDonPerugini]

Tend to agree. However, whoever owns the scalable standardized agent-to-agent internet protocol (agentic web + commerce, internet of agents) that will need to sit between consumer and business personal AI will become defensible infrastructure.

@SocontraInc

https://t.co/7QtNhDhH9D

[Dr Don Perugini @DrDonPerugini]

There have been lot of complaints about the usefulness of personal AI. The only use cases seem to be hotel reservations and travel, which do not occur often for most people.

I must agree.

I think these companies went headfirst into agentic commerce – using personal AI to make commercial transactions with businesses. This is not sticky and has no moat. I can easily replace Meta’s Muse for other personal AI like Instinct (who just raised $1B), OpenAI Dots or GrokBot.

Last year I built @SocontraInc agent protocol with the goal of helping startups build personal AI. A key component of the protocol was that it’s underpinned by a social network.

Why?

People are social, at home and at work. They like to connect and socialize. They work better when they work together or communicate and share knowledge (agents do as well). People typically connect with others often.

Using personal AI to facilitate this social interaction - my personal AI with their personal AI - is a great use case.

Some of the largest companies in the world are about social connections (Facebook, X, Slack, etc). Additionally, they have an unbreakable moat from the Network Effect – the more people that join (your friends, family, work colleagues), the harder it is to move or stop using it.

Personal AI companies that focus more on how personal AI can help connect us with our social networks first might be the winner in this race.

And once you facilitate the social connections and help people with automating tasks – guess what, those social events or tasks are likely to inherently involve commerce, e.g. buy and deliver a cake for the birthday party, make a reservation at the restaurant for the business meeting, organize travel with my friends or work colleagues for the conference, etc.

Talk more about this in the post in the comments.

[Dr Don Perugini @DrDonPerugini]

@HarryStebbings Exactly. We had http/html for websites (WWW),we need the same for agents, like @SocontraInc . MCP/APIs to connect agents is not standardized and scalable.

[Dr Don Perugini @DrDonPerugini]

You are describing the agentic web (internet of agents) and agentic commerce, where the internet will be filled with agents representing consumers and businesses (and websites). Agents acting on their owner's behalf will connect (directly or in groups), interact, collaborate and transact with each other.

That's why I built @SocontraInc - an open internet transactional protocol and social network for agents.

The social network effectively becomes a marketplace when agents interact and transact to buy and sell things.

[Dr Don Perugini @DrDonPerugini]

Agree, I talked a lot about this likely future outcone last year, and developed an agent protocol (@SocontraInc) to help developers expedite this future.

I think it's inevitable that everyone will have an AI assistant that will help them connect and communicate with others who also have AI assistant.

[Dr Don Perugini @DrDonPerugini]

Absolutely, personal agents interacting with and obtaining services directly from software apps is where AI is heading.

However, MCP and CLI as an interface is not scalable. Personal agents don't want to figure out how to use the specific tools from your app in order to achieve their intent.

A scalable model is personal agents just give you their intent, in say natural language, and an agent associated with your software app works out the specific tools and the their sequence needed to achieve that intent.

The internet of agents (or agentic web) then becomes scalable - agent's on the software app side communicate their general services to personal agents, and personal agents communicate their intent of what specific services they want. This is scalable.

@SocontraInc

[Dr Don Perugini @DrDonPerugini]

MCP (APIs) is not scalable and practical for agentic web - every app or web site will be different.

WWW was underpinned by the http protocol and html language.

Whats the equivalent for agentic web? Rather than MCP as an interface for agents, each app and website should use an agent as an interface. You then have an 'internet of agents' at both ends (the user agent, and app/web site agent).

A general agent protocol connects the two, like A2A or @SocontraInc - the http equivalent.

Rather than specfic tool calls, whats communicated is 'intent' in (mostly) natural language - the html equivalent. The app/web agent interprets the intent, calls its own complex combination of tool calls (which its best place to do) and responds satisfying that intent.

This is scalable and practical.

I have a PhD and spent 20 yrs in agent protocols (incl. Agentic web and commerce), which is why I built @SocontraInc .

But the industry is so far behind because as you said, no one is rethinking agentic web and just trying to use what we already have, which is why its not going anywhere.

[Dr Don Perugini @DrDonPerugini]

Thats why MCP is not scalable. Your product should be an agent and interface with other agents via an agent internet protocol like A2A or @SocontraInc .

Agents then communicate only thier intent (what they want) rather than having the complex task of working out (orchestrating) what API services they need to achieve their intent - the product agent is best placed to do that.

[Dr Don Perugini @DrDonPerugini]

Spot on. In the future all users and applications will be agents. This includes SaaS, websites, databases, etc. which will also have an agent front-end to interact with the other users' agents and provide services based on the user agent's intent. This will all need to be underpinned by an internet of agents or agentic web - the agent version of HTTP/HTML.

I have 20+ years experience in AI agents and agent protocols, and expected this to happen a lot sooner. Industry is a little slow, but I'm sure we will eventually get there. Discussed more in post below.

https://t.co/COcFN5VjiP

[Dr Don Perugini @DrDonPerugini]

@naval Yes. An 'agent protocol' for agents to interact, collaborate and transact over the open internet. Examples include A2A and @SocontraInc . Current APIs and MCP are not scalable.

[Dr Don Perugini @DrDonPerugini]

Have a PhD + 20 yrs experience in this area, and developed @SocontraInc agent protocol and social network for this problem.

AI agents have owners – people or companies – and act on their behalf. See agents as a digital twin of their owners (like a real estate agent acting on your behalf).

Therefore:
- The knowledge graph for each agent in the transaction represents the knowledge of the owner. Agents should transaction exactly like people/companies do in real life (and be consistent with the law).

- The owner defines success via its motives, goals, intentions or preferences communicated to the agent.

- The owner is responsible for the actions of the agent.

This means either:
(1) the owner should have final approval for any final transactions (as per your conclusion, best to have humans still in the loop); or

(2) trust the agent’s decision and take liability for its actions (over time, if the agent always successful, the owner may build trust to allow the agent to make decisions without oversight).

[Dr Don Perugini @DrDonPerugini]

Thats right. Need the right agent protocols for the open internet (eg. Agentic web) to enable agent-to-agent messaging, coordination and transactions, and infrastructure for agents to dynamically form public and private social groups (eg. Agent social network). Industry is way behind in this area, but I think will eventually get there.
@SocontraInc

[Dr Don Perugini @DrDonPerugini]

@gregisenberg 7, 13, 20 and 21 addressed by @SocontraInc agent protocol. 3 addressed by @CogFlowAI which uses cognitive science to capture tacit knowledge from experts to create cognitive skills/workflows (procedural memory). Great list.

[Dr Don Perugini @DrDonPerugini]

Agree. I would go further with regard to agent UX, and where I think the industry will (or should) go, and how easily the moat can be broken down for SaaS products (I have a PhD in AI agent protocols and thought about this a lot).

Opening up APIs is one step closer to agent UX or ‘agentic web’, but is still a barrier to agent integration/interoperability and its scalability. Agents still need to reason about the different APIs and tools for every call to say Brex. So moving from Brex to another provider still requires work to understand, implement, use, etc, a different set of APIs. How this looks like is (I hope this text illustration works):

Agent [Human/agent intent (natural language) -> LLM -> API calls] --> Comms [API calls] --> SaaS [API calls]

As you can see from the above diagram, what is communicated is API calls, which are different for each SaaS, are not standardized, not scalable and not suited to the agentic web. The internet (WWW for websites) became successful due to standard HTTP/HTML.

Where agent UX and agentic web should go is using natural language with human/agent intent as the interface and for communication. Rather than provide APIs, the SaaS has its own AI agent that receives requests/intent from other agents in natural language. The SaaS agent makes the API calls for its own SaaS product (which it is best placed to do so), and returns the result or performs the action. How agent UX looks like now is:

Agent [Human/agent intent (natural language)] --> Comms [Human/agent intent (natural language)]--> SaaS [Human/agent intent (natural language) -> LLM -> API calls]

In this model, natural language with human/agent intent becomes the new HTML that is communicated, is standardized and scalable, and makes it much easier for agents to access and swap between SaaS providers �� i.e. the intent needed for Brex should be exactly the same for another SaaS provider.

This also makes it easier for agent-to-agent interaction, transactions, etc. The model is the internet is full of agents that collaborate and transact via natural language, not half agents and half APIs. This is where we will see huge growth in agentic web.

For this to work you need an agent internet protocol like HTTP to underpin this new internet of agents (agentic web). Protocols like Google’s A2A or @SocontraInc are some examples.

[Dr Don Perugini @DrDonPerugini]

Great first step. However, imagine how awesome Salesforce would be if the UI was an AI agent instead of the API, MCP or CLI. The Salesforce AI Agent can receive simple requests from other agents - intentions of what they want via natural language - and return the relevant data, workflows, and tasks.

The Salesforce AI Agent interprets agents' intention and works out what APIs to call in the backend to achieve the intention. Developers won't need to know about the many API calls (tools), making it easier, quicker and more scalable for others to build AI agents and products with Salesforce.

This would also drive the idea of the ‘agentic web’ (or internet of agents). Of course, this would need to be supported by a general agent internet protocol (the HTTP for agents) like A2A or @SocontraInc.

Agentic web is coming - just slowly and in incremental steps.

[Dr Don Perugini @DrDonPerugini]

Great to see the AI agent industry finally identifying the need for an agentic web (internet of agents). Well, almost…

Over the last few weeks, industry leaders and investors had two separate insights:

(1)  All software products should have interfaces (API, CLI, MCP) to be relevant because future users of software will be AI agents and not humans. Google also released WebMCP so agents can interface with websites directly without “screen-scraping”.

(2)  There is a strong need for open payment protocols to enable agents to commercially transact with other agents (agentic commerce).

Addressing point 2 first, the internet was invented before e-commerce (carts and payments). Internet protocols like HTTP, with HTML, enabled the World Wide Web.

The same applies to agentic commerce and agentic web. We first need an agent internet protocol (like HTTP) to enable agents to connect (interface), collaborate and transact. We can then use payment protocols as a layer above this where transactions require payment.

Assuming that agentic commerce is underpinned by only payment protocols is placing the (shopping) cart before the horse.

Now to address point 1 above. I agree that software and websites need to interface with agents in a future where agents will dominate the internet over humans (agentic web), as agents will do more on our behalf.

However proposed interfaces (API, CLI, MCP, WebMCP) are not scalable. Each interface will have different services (tools) available, that need to be discovered and known, and reasoned by every agent each time they want to access the software or website (costly and timely).

Like with agentic commerce above, we need a standard agent internet protocol (like HTTP) for agents to connect and interface with each other. Importantly, we also need to be a standard language, like HTML is for the World Wide Web.

Since AI agents are intelligent, this language should be natural language, and the contents are the agents’ intentions of the services it wants.

In this way, rather than an agent trying to work out each service available for each software or website, it just communicates its intentions, and the software or website (which knows most about its own services) responds with the right services (tools, data, etc) that achieves the agent’s intention.

This also means that the interface for software and websites are not API, CLI etc., but AI agents that acts on its behalf, and interacts with other agents over the agentic web.

We are in the early stages, but the ‘type of protocols’ needed include Google A2A and @SocontraInc , and some others are emerging.