Using Agents in Production: Past Present and Future // Euro Beinat

A deep dive into how Prosus is deploying over 30,000 AI agents to create an 'AI Agentic Workforce'. The talk covers the transition from simple assistants to trusted senior colleagues, the internal tooling developed, and the crucial organizational strategies used to overcome adoption barriers and foster a bottom-up culture of innovation.

March 1, 2026 4 min read

Using Agents in Production: Past Present and Future // Euro Beinat — source: https://www.youtube.com/watch?v=tP8Xrwp-tNY

The Vision: An AI Agentic Workforce

At Prosus, a global e-commerce technology company, the goal is to have 30,000 AI agents in production by the end of March. This initiative is part of a dual strategy:

Reinventing E-commerce: Making customer-facing interactions more agentic, personalized, and ecosystemic. An example is iFood’s “ISO” agent, which guides users through a personalized shopping experience on WhatsApp.
Creating an AI Agentic Workforce: This is the primary focus, aiming to augment every employee with a suite of AI agents. The vision is to transition from current “AI assistants”—which are likened to eager but untrustworthy interns—to “senior agent colleagues.” These agents are designed to be domain-aware, reliable, and capable of executing complex, end-to-end tasks.

The expected benefits of this workforce are measured across four dimensions:

Productivity: Accomplishing more with the same resources.
Quality: Improving the output and performance on specific metrics, such as customer support scores.
Independence: Enabling employees to perform tasks without needing to rely on colleagues.
Agility: Empowering employees to work outside their primary comfort zones, such as a marketing professional building a small application without depending on an engineering team.

An Agentic Workflow in Practice

An agent is a tool trained to handle specific tasks by interacting with systems and other agents. A real-world example provided is handling a partner’s query about a recent drop in sales.

Traditional Process: A partner contacts a human account manager, who then asks a data analyst to pull a report. This involves back-and-forth communication and can take anywhere from an hour to a day.
Agent-based Process: The partner’s query, sent via WhatsApp, is received by an AI Account Manager agent. This agent tasks an AI Data Analyst agent, which is capable of querying data sources like DataBricks or Tableau. The data analyst agent retrieves and packages the information, returns it to the account manager agent, which then formulates and delivers an instantaneous answer to the partner, while also notifying the human manager.

This new workflow is not only instantaneous but also provides universal coverage, supporting all partners instead of a select few.

Building and Scaling Agents Internally

Prosus identified two classes of use cases:

Well-structured, frequent use cases (e.g., coding, sales) with large addressable markets, often served by existing commercial tools.
A long tail of sparse, specific use cases unique to individuals or small teams. While each is small, their cumulative value is massive.

To address this long tail, Prosus developed “Tokan,” an in-house AI agent builder. The decision to build versus buy was driven by needs for security, privacy, cost-effectiveness, and the ability to create custom integrations with internal systems.

Agents built with Tokan are classified by a seniority model, which correlates with the number of tools and integrations they can access:

Interns: Can work with documents.
Juniors: Have read-access to systems.
Intermediates: Can write data back to systems, requiring a higher level of trust.
Seniors: Can orchestrate other agents to perform more complex tasks.

The Key Challenge: Fostering Adoption

The most significant learning was that after a certain point, AI adoption is not a technical problem but an organizational and cultural one. Initially, the creation of agents grew linearly, driven by new features and integrations. The challenge was to trigger exponential growth.

The primary barrier was a perception among non-technical employees that building agents was a complex task reserved for engineers. To overcome this, the team focused on “making agents uncool”—demystifying the process and showing that anyone could build them.

This was achieved through several initiatives:

Hands-on Demonstrations: Working directly with sales, HR, and marketing teams to quickly build agents that solve their specific problems, demonstrating the simplicity and iterative nature of the process.
Upskilling at Scale: Providing broad training on the tools and features available.
Fostering an Agent-First Mindset: Encouraging employees to think of an agent as a potential solution for any “job to be done.”
Internal Competitions: Launching programs like “Prosus Got Talent,” an internal, Shark Tank-inspired competition where teams from across the business build agents to solve real challenges. This ignites innovation, gives recognition, and helps build the “organizational muscle for AI.”

Core Learnings and Conclusion

The journey to deploying 30,000 agents yielded three crucial insights:

Adoption is an Organizational Problem: Beyond providing the right tools, the focus must shift to culture and empowering people.
Embrace Bottom-Up Discovery: The best way to find valuable use cases is to enable everyone to experiment. This process of collective discovery at scale becomes a competitive advantage.
Surface Domain Knowledge: The effectiveness of agents is directly tied to the ability to surface and encode implicit domain knowledge into their workflows.