# A New Kind of Marketplace

Donné Stevenson, Prosus & Pedro Chaves, OLX Group | MLOps Community | 51:27

Source: https://www.youtube.com/watch?v=q9e2e5Y8Q0k
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/a-new-kind-of-marketplace
Published: 2026-04-20
Tags: agents, ecommerce, guardrails, multi-agent

## TL;DR
- A useful agent experience starts by changing the user's task, then earns trust through small, controlled releases.
- Marketplace agents need interfaces that combine chat, shortcuts, generated components, and explicit confirmation before taking actions.
- The next marketplace layer may be infrastructure that lets agents discover, negotiate with, and transact with other agents.

## Summary
Donné Stevenson and Pedro Chaves discuss how agents could change buying and selling across real estate, automotive, and general classified marketplaces. Pedro describes a real-estate experience that treats house hunting as a lifestyle decision. An agent gathers preferences about neighborhoods, amenities, schools, safety, and commuting, then combines chat with clickable components and enriched recommendations. For dealers, Donné describes an early agent that helps interpret marketplace data through chat and shortcut actions. The discussion returns repeatedly to trust. Users need to understand what an agent can do, see what will happen, and have a way to undo actions. The longer-term idea is a marketplace where buyer and seller agents discover each other, negotiate within stated constraints, and use an escrow or payment layer. Pedro argues that companies should build the infrastructure, or "harness," around agents rather than compete only on the agent itself. Physical delivery may connect to lockers, couriers, robots, or drones.

## Key ideas
### A new marketplace experience starts with a different definition of the user's task
[01:04](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=64s)
Pedro says OLX operates across motors, real estate, and general classifieds, with many professional sellers such as dealerships and real-estate agencies. The hard product problem is defining the disruptive experience before building it. In real estate, the team moved beyond filters, maps, budget, and property attributes. Looking for a house is treated as a lifestyle event that includes neighborhoods, parks, services, schools, safety, and commute time. An agent can extract these preferences, model the user's intent, and match them against enriched listings.

### The first agent interface should mix conversation with visible controls
[05:21](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=321s)
Pedro rejects the assumption that house hunting should be pure chat. The proposed experience combines conversation with pre-populated components, clickable choices, navigation, and an enriched map or listing page. Donné describes a similar pattern for automotive dealers. Their agent presents data insights alongside shortcut buttons, some powered by AI and some implemented as ordinary workflows. These controls give users a clearer starting point than an empty chat box and help them understand what the agent can do.

### Trust grows through staged releases and limited actions
[07:46](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=466s)
The teams validate new experiences through research, prototypes, internal rollouts, real-estate experts, an invite-only public beta, and then a wider release. Each stage provides explicit feedback and behavioral signals. Donné says dealers initially avoided a shortcut because its label made them fear that it would act on every advertisement. The team changed the progression so the agent first shows relevant items, then asks whether the user wants an action taken. Pedro and Donné connect this gradual expansion of permissions to the trust required for agents with access to financial or business data.

### Agent products need a learning curve that does not punish mistakes
[16:23](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=983s)
The speakers say users often do not know what to ask a raw chat interface, especially when the agent has limited functionality. They may ask nothing or submit an oversized request. Donné argues that preset actions can teach users what is possible and lead them into follow-up questions. They also discuss a control-Z-style undo function. If users know an action can be rolled back, they may be more willing to let an agent act. Pedro suggests interactive components that appear when relevant, allowing a user to approve an action instead of replying to a vague request with yes.

### Classified-marketplace agents could search, negotiate, and schedule for buyers
[24:33](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=1473s)
Pedro describes a buyer who gives an agent a target item, budget, flexibility, and time constraints. The agent searches listings, keeps monitoring them, and negotiates with sellers when a suitable item appears. It could also handle the back-and-forth over meeting times. The example comes from buying a child's scooter on OLX, where arranging the meeting and obtaining cash created unnecessary work. An agent could reduce that friction if it had access to payment methods and a trusted transaction process.

### An agent-to-agent marketplace needs a trusted middle layer
[28:57](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=1737s)
Pedro imagines autonomous buyer and seller agents interacting around second-hand goods. He says the missing piece is a marketplace harness, or an infrastructure layer that exposes listings, lets agents discover one another, communicates protocols and preferences, and supports negotiation. An escrow function could hold money and reduce fraud. Donné asks why agents could not communicate directly, but Pedro says they still need a discovery and mapping layer because different assistants and technical approaches will coexist. He sees this layer as a possible marketplace product.

### The marketplace should build the harness around agents rather than another general assistant
[32:23](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=1943s)
Pedro argues that large technology companies already provide increasingly capable assistants, and their capabilities can be extended with tools or MCP. A marketplace company can focus on the environment those agents use: posting items, uploading descriptions and images, discovering available agents, and negotiating transactions. He calls this the exoskeleton for an agent marketplace. The company would provide access to supply and demand without needing to own the agent that users choose.

### Physical logistics may connect agents to lockers and existing delivery systems
[43:53](https://www.youtube.com/watch?v=q9e2e5Y8Q0k&t=2633s)
The conversation extends agent-mediated buying into the physical world. Smart lockers could remove the need for buyers and sellers to coordinate a meeting and could reduce the risks of direct contact. A seller might photograph unwanted items, receive offers from interested buyer agents, confirm a sale, and place the goods in a locker or hand them to a courier. Donné mentions autonomous delivery robots and drones, while Pedro says OLX should focus on enabling agent interactions and connect to logistics providers rather than solve every last-mile problem itself.

## Notable quotes
- Pedro Chaves: "The first thing that was really required and is still required is to actually understand what's this new experience." (01:28)
- Donné Stevenson: "Functionality doesn't matter until your UX is solved." (18:15)
- Pedro Chaves: "What we really need is the infrastructure with which they interact." (32:44)
- Pedro Chaves: "How do we unlock this item circulation?" (49:00)

## Tools & references mentioned
- OLX Group
- Prosus
- MLOps Community
- Cursor
- Gemini
- Claude
- OpenAI
- Shopify
- Stripe
- Tencent
- WhatsApp
- MCP
- Google Pay
- Apple Wallet

## Who should watch
- You are designing an agent for a marketplace and need to decide how much of the experience should be chat, controls, or generated UI.
- Your product handles valuable or sensitive actions, and you need a practical approach to permissions, staged rollout, trust, and undo.
- You are exploring agent-to-agent commerce and want to think through discovery, escrow, negotiation, and the boundary between marketplace infrastructure and logistics.

## Related talks

- [The Agent Exchange: Practitioner Insights](https://mlopstalks.com/talks/the-agent-exchange-practitioner-insights) (Dmitri Jarnikov & Chiara Caratelli, Prosus Group & Steven Vester, OLX, 48:18)
- [AI Agents Are Revolutionizing E-Commerce](https://mlopstalks.com/talks/ai-agents-are-revolutionizing-e-commerce) (Nishikant Dhanuka & Beatriz Ferreira, OLX, 29:58)
- [The Agent Landscape - Lessons Learned Putting Agents Into Production](https://mlopstalks.com/talks/the-agent-landscape-lessons-learned-putting-agents-into-production) (Paul van der Boor & Floris Fok, Prosus Group, 1:08:41)
- [Before Building AI Agents Watch These Hard Earned Lessons](https://mlopstalks.com/talks/before-building-ai-agents-watch-these-hard-earned-lessons) (Nishikant Dhanuka, Prosus Group, 52:37)
- [Web Agents: The Cutting Edge of AI is Here?](https://mlopstalks.com/talks/web-agents-the-cutting-edge-of-ai-is-here) (Paul van der Boor & Chiara Caratelli, Prosus Group, 45:53)