# The Agent Exchange: Practitioner Insights

Dmitri Jarnikov & Chiara Caratelli, Prosus Group & Steven Vester, OLX | Agents in Production Series | 48:18
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=2s_N4EFcIFg
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/the-agent-exchange-practitioner-insights
Published: 2025-03-03
Tags: agents, ecommerce, multi-agent, product-strategy

## TL;DR
- Dmitri Jarnikov argues that useful agents will specialize in reasoning while combining specialized execution with a generic fallback.
- Steven Vester expects adoption to begin with narrow, trusted shopping tasks before users move to generic agents.
- The panel expects e-commerce platforms to resist third-party agents at first, then cooperate with them when agents become common.

## Summary
This episode examines two questions about AI agents in e-commerce: whether generic or specialized agents will win, and what happens to companies that make shopping agents. Dmitri Jarnikov separates reasoning from execution and argues for agents that are specialized where precision matters, while still falling back to generic browser use. Chiara Caratelli describes a similar mix, with specialized agents handling familiar platforms and generic agents covering unfamiliar sites. Steven Vester expects users to start with small purchases, build trust through successful experiences, and gradually delegate more complex tasks. The panel also discusses the conflict between an agent seeking the best deal and a platform seeking revenue from recommendations, advertising, and customer relationships. Steven expects platforms to block agents before opening APIs and data access. Dmitri compares this shift with Spotify and music labels. The conversation also raises selling agents, though Steven is unsure what they add beyond existing inventory tools.

## Key ideas
### Useful agents need specialization in both reasoning and execution
[02:37](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=157s)
Dmitri Jarnikov separates agents into generic systems that can operate a computer or browse the web, and highly specialized systems that perform one task on one service. Specialized agents are more precise, but building one for every shopping site does not scale. Generic agents scale better, yet Dmitri says they handle only short sequences reliably, with common examples reaching about six steps before failing. He evaluates agents along two dimensions, reasoning and execution, and wants specialization in each where it helps. His preferred design combines specialized reasoning with specialized execution, plus a generic fallback. He uses food ordering as an example because the agent must understand preferences, build a food order, and work across several services.

### Agent systems can dynamically combine generic and specialized components
[07:39](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=459s)
Chiara Caratelli says she began with a generic web agent, moved toward specialized agents, and ended up favoring a balance. Agents do not need a fixed workflow. A generic manager can call specialized agents, while a specialized shopping agent can use a generic browser on smaller sites it has not seen before. She says this mixed approach is easier with agents than with earlier machine learning systems because the system can choose its path dynamically. Specialized agents have better success rates inside their intended scenarios and are faster for real-time interaction. Generic browsing can still work for slower tasks such as deep research, where the user cares more about reaching an answer than about response time.

### Users will adopt agents through small, successful tasks
[10:51](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=651s)
Steven Vester frames adoption through the example of a human assistant who first receives simple jobs, such as getting coffee, before being trusted with a report or a research plan. Users will likely begin with narrow purchases, such as movie tickets, then delegate larger purchases after seeing a record of success. Steven distinguishes actual capabilities from users' perceived capabilities. Generic agents currently make errors and are not good enough for important shopping tasks, so he expects specialized agents to produce the first successful experiences. Over time, better models should improve generic agents, and users may treat them like online payments. His prediction is that specialization will lead early adoption, while generic agents will become normal as their capabilities and users' trust improve.

### Platform incentives may conflict with an agent's search for the best deal
[20:38](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=1238s)
The panel points out that shopping agents could remove the random browsing and recommendations that generate extra purchases for e-commerce platforms. Steven says an ideal agent would define the user's criteria and search an entire inventory, potentially across several platforms. That requires better access than clicking through millions of listings. Platforms might expose searchable inventory, APIs, or embeddings in a vector database, while Chiara suggests platforms could rank results with their own recommendation models and include paid content. The user may benefit from broader search, but the platform may prefer to show items that make more money. Chiara says users will expect personalization and the best deal, and trust could disappear quickly if paid placement distorts the agent's answers.

### E-commerce platforms are likely to resist agents before cooperating
[26:38](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=1598s)
Steven expects marketplaces to begin defensively because third-party agents can weaken brand loyalty and make customer behavior harder to measure. He compares the problem with price comparison sites, which can disintermediate classified platforms. Possible defensive measures include bot detection, blocking data-center traffic, and honeypots that feed scrapers misleading data. Steven doubts this approach will last. Once agents become common, platforms will need to make their inventory and APIs easier for agents to use or risk losing users to competitors. That change would also weaken business models based on impressions and views. Dmitri agrees that blocking agents creates a cat-and-mouse problem because stronger blocking can also make the site worse for real users.

### Agent companies and platforms may form a relationship like Spotify and music labels
[33:00](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=1980s)
Dmitri compares shopping agents with Spotify's relationship with music labels. Spotify focused on serving listeners, while labels controlled artist supply and needed usage information to make decisions. He expects a similar bargain in e-commerce. A generic agent could focus on the user's experience, while a marketplace could provide APIs, an agent-optimized site, or catalog access. In return, the platform would retain customer data and continue working with suppliers. Dmitri says a platform might build its own shopping agent, but a single-platform agent would lose much of the value of comparing products across competitors. A generic agent can provide a marketplace abstraction across OLX, Allegro, Amazon, and other services, though building and maintaining that scope would be difficult.

### Selling agents lack the shared infrastructure that shopping agents need
[42:36](https://www.youtube.com/watch?v=2s_N4EFcIFg&t=2556s)
Dmitri says the discussion has focused on agents that buy, while agents that sell are less developed. A selling agent could help a person sell an unwanted item, a manufacturer offer products, or a service provider find customers. He thinks the missing piece is infrastructure connecting selling agents with shopping agents. Steven is less certain that selling agents add much for car dealers and real estate agents because inventory management systems already distribute listings across several platforms through APIs. Agents might automate replies or expand distribution, but professional sellers often want to stay close to potential customers because that is where their sales skill matters. The panel leaves this as an open subject rather than predicting a clear outcome.

## Notable quotes
- Dmitri Jarnikov: "The future is for agents that are reasonably specialized in reasoning and have a combination of specialization and execution with a possibility to fall back on generic execution pattern." (06:46)
- Steven Vester: "Today generic agents are just not good enough to really do stuff." (13:23)
- Steven Vester: "I think today it needs to be specialized because generic is not good enough." (15:39)
- Steven Vester: "I think that they will start being very defensive in this in the beginning and then we'll move to embrace it." (27:01)
- Dmitri Jarnikov: "There is so much need also for selling agent." (42:36)

## Tools & references mentioned
- OLX Magic
- OpenAI Operator
- DeepSeek
- Gemini
- Spotify
- Amazon
- Google Pay
- PayPal
- Prosus AI

## Who should watch
- You are building a shopping or web agent and need to decide where specialization belongs in reasoning, tool use, and browser execution.
- You work on an e-commerce marketplace and need to think through bot defenses, agent access, customer data, and future monetization.
- You want a grounded discussion of user trust, recommendation incentives, and the possible role of selling agents.

## Related talks

- [A New Kind of Marketplace](https://mlopstalks.com/talks/a-new-kind-of-marketplace) (Donné Stevenson, Prosus & Pedro Chaves, OLX Group, 51:27)
- [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)
- [AI Agents Are Revolutionizing E-Commerce](https://mlopstalks.com/talks/ai-agents-are-revolutionizing-e-commerce) (Nishikant Dhanuka & Beatriz Ferreira, OLX, 29:58)
- [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)
- [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)