Hari Vasudev, CTO of Walmart’s US business, says that he expects agents to drastically change the way companies like Walmart engage online. He points to them building agents of their own to interact with the consumers’ agents to provide recommendations or additional product information, while the consumer agents could provide the retailer agents with information about preferences and so on. This coming world of consumer agents talking directly with retailer agents fascinates me.
Agent Commerce
A simple model of this future commerce looks like this: The buyers’ agent asks the seller for a connection to their agent. This is going beyond shopping bots capable of hopping between online retailers searching for the best prices, deals and promotions for their users and on to rich conversations and exchanges of data to enhance the connection between consumer and retailers.
(Salesforce released its Agentforce platform, which lets retailers build AI agents that can interact with their systems, last year and their CEO Mark Benioff’s stated ambition for 2025 is to have thousands of customers live on the platform with “billions” of consumers interacting with agents.)
What if the seller does not have an agent? Well, if the seller does not have an agent, then the buyer’s agent will instead look for APIs that it can use. At the Merchant Payments Ecosystem conference in Berlin this year, Simon Redfern from the Open Banking Project gave a live demo of their Opey agent using a database of API endpoints to assemble queries and obtain the data that it needed to complete a task.
(They describe Opey as an emergent property of a formally documented and tagged API system, pre-processing software and AI which can synthesise the abilities of those APIs. You can ask Opey questions such as “What endpoints can I use to get data about ATMs?”)
What if the seller has no APIs? Well, if there are no APIs, then the agent will just use a web browser to get what it wants. There has been a lot of coverage the Manus ai agent recently, and it provides a window into the future. Give it a task, such as booking tickets and hotels for a conference, and the agent will spin up a version of itself to browse the web and go for it. The Manus ai system sits on existing models that can interact with the internet and perform a sequence of tasks without deferring to a human user for permission. What makes Manus rather interesting is that it has been released by its creators to spark a new era of experimentation out in the real world.
Not everyone is happy about the paradigm shift from e-commerce to agentic commerce (a-commerce). In a meeting between DoorDash and OpenAI (owners of the Operator, a bot that uses a web browsers), an executive from the former noted an obvious implication: If only agents visit the DoorDash site, that wouldn’t be so great for DoorDash’s business. Well, it wouldn’t be good for their current business where they show ads to people. Maybe they could find out what bots like and advertise to them instead through some verifiable attestations to response times, server uptimes, hallucinations or whatever.
This is fascinating to speculate about, especially as the likes of Walmart get involved: how will this shift from e-commerce to a-commerce work in practice? There are many issues to work through, but I think there are two to highlight for the moment: there are discovery and reputation issues that need to be thought through.
Agent Discovery
The first issue is discovery. In practice, how will my agent find your agent, and when it finds it, how does it know whether to use it or not? One might imagine that there will be exploration in different directions here, but for now we might think of systems of decentralized directories or registries that would maintain a list of active agents and their capabilities, some set of standard communication protocols to allow agents to broadcast their needs or offers and perhaps even the equivalent to social media where agents could form connections based on past interactions, which can help in future discoveries.
(I wonder if a class of influencer bots will emerge to nudge toward certain other bots.)
Once agents can discover each other and trust each other, then it is a natural step for them to incentive and reward each other through payments. This will mean that agents will be invited into the payment system and will transact and negotiate dynamically on behalf of individuals, businesses and even with other AI agents — creating a radically different commercial landscape.
David Minarsch of Valory, the core contributor to the OLAS open source software stack that uses smart contacts to connect agent buyers to agent sellers in an Agent Bazaar (there are some 550 agents in the bazaar already), told me that he thinks the use of stablecoins will really open up a-commerce by enabling precisely this kind of dynamic environment, and I think he makes a good point.
Agent Reputation
Payments are part of the new landscape but another part is identity. In the case of OLAS, agents are identified using wallet addresses for now, but it is easy to see that a more universal (and accessible) identification system will be required in the future. Once agents have identities, they can have credentials attached to them (eg, “this is an authentic NYSE price bot” or whatever) and once they have credentials then they can, over time, address the second issue: reputation.
This seems rather important to me. People are not very good at discovering and maintaining reputation: when you last ordered some take out food, did you check that the food really came from the advertised restaurant, that the hygiene checks were up to date and that the burrito reviews were real and planted by restaurant sock puppets? Of course not. And while reputation probably doesn’t matter too much for burritos, it does matter a lot to business.
In the agent economy, where they may be millions of agents to choose from for any particular, task, reputation is something that is discoverable and (when implemented using the appropriate cryptography) not game-able. This will mean not only a more efficient marketplace but a safer one.
Read the full article here
