Rohlik Group: Building the agent-driven grocery company

Who they are

Rohlik was one of the first unicorns founded in the Czech Republic, and after ten years on the market it remains one of the country's most loved brands. From the outside it looks like a grocery company.

Under the hood, it runs as a technology company. Rohlik's culture is built around a single metric — the perfect order — and hitting it at scale requires AI and robotics at every step: an AI shopping assistant that takes customers from browsing to checkout, automated warehouses where robots store and retrieve products across a grid, demand forecasting at SKU-level granularity, and routing systems that guarantee a courier arrives within a 15-minute window across an entire city.

Nearly all of that software is built in-house. Rohlik's warehouse control system, Arcos, decides which bin to pick, how to pack products into bins, and how to prioritize orders, and it delivers roughly 10% more efficiency than comparable systems on the market. Arcos sits inside a broader operations brain — a stack of models that continuously fuses demand signals, warehouse execution, and courier data to steer fulfillment and last-mile delivery in real time, re-prioritizing picks and routes as conditions shift minute to minute. The demand forecasting has to be precise enough to order fruit that perishes in twelve hours. The company operates twelve fulfillment centers across five countries, so a software bug can stop physical operations — the cost of an error is far higher than for a typical software company.

The ambition

Rohlik's business is built on velocity. The company continues to expand across Europe, and also package its technology as a platform, Veloq, that other grocers can use to deliver the same customer experience. It is also pushing toward faster delivery through networks of small fulfillment centers spread across cities — a model that demands an order of magnitude more from its orchestration software, from demand prediction at a very granular level to digital-twin monitoring that lets a small central team run dozens of sites.

Reaching those goals requires shipping software much faster than the company's size would traditionally allow.

"In order to achieve the ambitious plans we have, we need to be able to move really fast. If we want to maintain this position of being at the forefront of innovation in online grocery, we need to be able to essentially multiply our engineering output."

— Michal, Group CTO

The physical nature of the business raises the bar further. When Rohlik decided to integrate solutions like AutoStore robotics into its warehouses, they found out integration software needed to be entirely on Rohlik's side in order to be able to evolve operations fast. Then they executed within 8 months a kind of project that takes most companies two to three years.

Turning to agents

Rohlik began its agent journey about a year ago, starting from zero agent-generated code. The team's early experience with Devin was mixed.

"When we set up Devin a year ago, it was kind of buggy. It didn't manage the full SDLC, so you still had to give it a lot of hints. But after a year it's able to do everything. You give it a task, you ask it to test it, you ask it to check the production logs, check the metrics, suggest improvements. If your company pays attention to the tooling setup, Devin can just handle it."

— Michael, tech lead for last-mile logistics

What made the difference was the criteria Rohlik cared about: an agent that could complete work end-to-end, test its own output, run in the background in parallel, and plug into the company's systems. Rohlik invested in the foundations to make that possible — over 50 MCP integrations, internal and external, built on the principle that any new tool should be agent-accessible from day one, plus a semantic layer over its Snowflake data warehouse and a knowledge base that gives agents the context a new colleague would need.

How Rohlik works now

Today Rohlik's engineering organization is, in Michal's words, "agent-mostly." About 90% of code is agent-generated, and agents operate in every step of the SDLC: planning, code generation, code review, and incident response.

Work flows into Devin from wherever it appears. Engineers spin up sessions directly from Slack conversations about a bug or feature, or take a PRD from Linear into Devin and work it through to a PR. When a product specification lands, Devin is unleashed on it first — if it can't implement the spec, it feeds back to the product team what needs to be clarified before an engineer ever touches it. Automated workflows triage incoming bugs. Devin is part of automatic incident response, and scheduled automations run code improvement and error-fixing flows in the background.

One of the most advanced examples is Maia, Rohlik's AI shopping assistant, which handles 90% of customer emails and chats and 30% of voice communication, and now supports agentic payments — customers can go from browsing products to checkout entirely through the assistant. Most of Maia is agent-written, and it improves itself: when a customer reports a bad experience, a PR is automatically created from that feedback to fix the issue.

Adoption reaches well beyond engineering. About 30% of Rohlik's Devin usage is data analysis, with business users across the company asking questions directly against the data warehouse.

"Data analytics was a bottleneck before. If you wanted to ask a question, you needed to ask someone to prepare the analysis. We are now at the point where that's no longer the bottleneck — the bottleneck is the ideas themselves."

— Michal, Group CTO

Product managers submit PRs directly. In the warehouse organization, agents handle exception management — preparing vendor claims from quality issues on inbound deliveries, contacting vendors, and following up on invoices — and refine labor planning by reasoning about factors like expected no-shows before a long weekend.

The organizational transformation

Rohlik rebuilt the organization around agents. Teams got smaller, and engineers became full-stack by requirement, because the biggest gains come from removing handoffs between people. Engineers are given a goal and a metric and have free hands in how they reach it, operating as product engineers who make product and design decisions.

"A year ago it was, we wish you were full-stack, give it a try. Now it's a requirement, because the agents just got a lot better."

— Michael, tech lead for last-mile logistics

The daily rhythm changed with it. "You come to the office, you check what's needed from you, and you offload it to an agent," Michael says. His team of twelve people runs, in his words, "a couple hundred agents." Because Devin works in the background from Slack and the web app, engineers create PRs from public transport and keep sessions running through the day. The skill mix shifted too: where engineers once spent most of their time writing code, Rohlik's best engineers now spend 80% of their time reviewing it, and the team uses Devin Review to push on that new bottleneck.

Results

Since November, Rohlik has doubled engineering output, and Michal believes the team is just scratching the surface. The AutoStore warehouse robotics integration went from design to running product in eight months, a project that takes most of the industry two to three years. Prototyping collapsed from a month of development to a day, which changes how the company makes decisions: teams test fifteen variants of an idea, ship prototypes to users for feedback, and only then invest engineering effort in the winner. In a business where systems failure stops warehouses in five countries, faster shipping has come with the safety gates intact — prototyping is kept separate from production-grade code, with process and business-logic checks alongside technical ones.

The vision

Rohlik is working toward what Michal calls hands-off engineering: work flowing from an idea to code generation through review to merge without a human in the loop, proven out one portion of the codebase at a time.

"For the next three to six months, we really need to get used to the fact that we will not see the code that gets into the codebase. We need to rely on the systems that we've built and the context that we've created. The concept of hands-off engineering will be a very real future."

— Michal, Group CTO

The warehouse organization shares the ambition: within a few years, Rohlik wants fulfillment operations that need minimal skilled management to run, monitored through real-time digital twins from a central point. And the end state reaches past engineering entirely.

"The end game should be that a business stakeholder — a manager in finance, operations, supply chain — will be able to implement changes themselves. I think — this is the new process I want to implement, and a partner like Devin will just do it. I can see that is the future."

— Dalibor Pátek, product lead for warehouse automation