# Using Agents in Production: Past Present and Future

Euro Beinat, Prosus | Agents in Production 2026 | 23:12

Source: https://www.youtube.com/watch?v=tP8Xrwp-tNY
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/using-agents-in-production-past-present-and-future
Published: 2026-03-01
Tags: agents, evals, team-adoption, tool-use

## TL;DR
- Prosus expects to have 30,000 agents in production across the group by the end of March, with agents supporting both customer-facing products and employees.
- Agent adoption becomes an organizational problem after the basic technical tools are available, so Prosus uses demonstrations, upskilling, integrations, and competitions to help employees build agents.
- Agents become more capable as they gain tools and system integrations, but their usefulness depends on encoding the domain knowledge that people use in their work.

## Summary
Euro Beinat describes how Prosus is deploying agents across its e-commerce businesses and internal workforce. The company expects 30,000 agents in production by the end of March. Customer-facing examples include an agent that supports food-delivery shopping and a group of agents that answer sellers' questions about changes in sales. Internally, agents range from simple document assistants to senior agents that coordinate other agents and access several systems. Prosus built its own agent builder, Tokan, because it needed control over safety, privacy, integrations, and cost. Beinat says the hardest part has shifted from building features to helping employees adopt the tools. Prosus gets non-engineering teams to build agents, provides training, and runs competitions around real business problems. He also describes four ways to evaluate an agent: productivity, quality, agility, and independence. The talk is practical about the limits of large-scale deployment, especially the difficulty of evaluating agents and encoding implicit domain knowledge.

## Key ideas
### Prosus is using agents to change customer experiences and employee work
[00:04](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=4s)
Prosus expects to have exactly 30,000 agents in production across the group by the end of March. Beinat describes two uses. Customer-facing agents can make e-commerce more personal by understanding intent and guiding a purchase. Internal agents form an AI workforce that helps employees do more, improve quality, work more independently, and take on tasks outside their usual comfort zone. Beinat expects employees may have one, two, ten, or even twenty agents helping them.

### A group of agents can answer a seller's business question without the usual handoffs
[05:07](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=307s)
Beinat gives the example of a seller asking why sales have recently dropped. In the old process, a partner manager would ask a data analyst for help and then return the answer to the seller. In the agent-based process, an AI account manager asks an AI data analyst for information. The analyst can query Databricks, Tableau, and other data sources, package the result, and send it back for review. Beinat says the answer can arrive instantly and cover all restaurant partners rather than only 20% or 30% of them.

### Prosus separates common workflows from thousands of small internal use cases
[07:58](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=478s)
Beinat divides agent applications into two groups. The first covers structured, frequent workflows such as technology, coding, customer support, and sales. These areas are similar across organizations, so dedicated businesses already support them. The second group contains sparse, highly specific tasks that may belong to one person or a small team. Prosus did not find a general tool for this long tail, so it built its own tools and agents to capture the value spread across many small applications.

### Tokan gives Prosus control over safety, integrations, and cost
[09:12](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=552s)
Prosus built Tokan, an internal agent builder that is not sold externally. Beinat says the company chose to build it for safety and privacy, for the ability to optimize for its own use cases, and for connections to its own integrations and MCPs. Cost was another reason. Tokan began in 2019, was released shortly before ChatGPT, became agent-based in 2024, and began helping employees create agents about a year before the talk. Prosus released the tool to colleagues in December.

### Agent seniority increases with tools and system access
[17:22](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=1042s)
Prosus classifies agents from interns to seniors. Intern agents work with documents. Junior agents can access systems through credentials. Intermediate agents can also write data back to databases, which requires more trust. Senior agents orchestrate other agents and handle more complex work. Beinat connects seniority with the number of available tools and integrations. These integrations can include email, calendars, systems of record, help desks, company-specific systems, and the web.

### The useful agent is tied to a specific job and measurable outcome
[12:48](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=768s)
Beinat describes a restaurant account executive that gathers information from several systems, creates a web page, and gives account managers material for restaurant meetings. About 200 colleagues use it, and Beinat says it does work equivalent to about 30 full-time employees. A data analyst agent lets people ask questions in ordinary language and access databases without writing SQL. Beinat's newsletter agent does not save him practical time, but it gives him broader information coverage and lets him personalize what he reads.

### Agent adoption grows when employees see that building one is ordinary work
[16:06](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=966s)
Prosus saw agent creation grow continuously but not exponentially until it addressed organizational barriers. Many employees believed that only engineers could create agents. The company worked directly with sales, HR, and other teams to build examples quickly and show that people could iterate on an agent themselves. Beinat says Prosus had to make agents 'uncool' in the sense that they should feel like ordinary tools for getting work done. Training, new integrations, and competitions then helped employees develop the habit of looking for an agent-based solution to a job.

### Prosus evaluates agents through productivity, quality, agility, and independence
[22:13](https://www.youtube.com/watch?v=tP8Xrwp-tNY&t=1333s)
When asked how the company evaluates thousands of agents, Beinat says there is an evaluation framework that is still being developed. Some agents can be tested with A/B comparisons against work done without the agent. Others require before-and-after comparisons or feedback from users. He names four evaluation dimensions: productivity, quality, agility, and independence. The exact implementation depends on the use case, so there is no single evaluation method for every agent.

## Notable quotes
- "The focus in terms of technology at this moment that we have this big technology shift is mainly on two things, in using AI agents and large commerce models." (01:07)
- "AI adoption is never a technical problem, is always an organizational problem." (16:54)
- "We had to make agents uncool." (17:35)
- "This is a bottom-up process. It is a process of collective discovery." (20:24)
- "These tools work when we can surface and encode domain knowledge." (20:42)

## Tools & references mentioned
- Prosus
- WhatsApp
- Databricks
- Tableau
- Tokan
- ChatGPT
- MCPs
- Shark Tank
- SQL

## Who should watch
- You are responsible for rolling out agents beyond an engineering team and need practical ways to get sales, HR, operations, or other groups to build them.
- Your organization is building many internal agents and you need a simple model for capability levels, system access, and evaluation.
- You want examples of how agents can support data analysis, account management, and information gathering in a large e-commerce company.

## Related talks

- [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)
- [We're Using AI Agents at Work (and it's amazing)](https://mlopstalks.com/talks/were-using-ai-agents-at-work-and-its-amazing) (Euro Beinat, Prosus Group, 27:37)
- [AI Agents: The Future of Productivity, or Just a Fad?](https://mlopstalks.com/talks/ai-agents-the-future-of-productivity-or-just-a-fad) (Sam Partee, Arcade AI, 35:18)
- [Why Enterprise AI is Taking So Long to Implement](https://mlopstalks.com/talks/why-enterprise-ai-is-taking-so-long-to-implement) (Paul van der Boor & Sean Kenny, Prosus Group, 1:05:01)
- [How to build agents that take ACTION](https://mlopstalks.com/talks/how-to-build-agents-that-take-action) (Alex Salazar, Arcade, 28:59)