How Delivery Hero's agent merges 100+ pull requests a day with Claude

Centralizing engineering across a global network of teams

Since its founding in 2011, Delivery Hero has brought local delivery leaders into a global group of brands that now processes over 11 million orders daily. For years, each of these local brands ran with near-total autonomy over technical decisions. That approach worked during hypergrowth, but as the company began centralizing around 2021, it created a problem: every team had different build systems, different infrastructure setups, and deeply held preferences about all of it. Some teams still had traditional structures with strict distinctions between developers, testers, and system administrators. Others had adopted cross-functional teams that owned, tested, and operated everything they built.

"There's always a certain history and attachment to whatever engineers have at the moment," said Rodrigue Schäfer, Vice President of Platform at Delivery Hero. "Convincing everybody that a global developer platform might be a good idea took around a year and a half."

The challenge extended to AI tooling. Starting in mid-2024, Delivery Hero entered an experimentation phase, allowing different teams to pilot different solutions. The results were promising across the board, but the decentralized approach made it hard to measure success consistently or push for adoption at scale. Many engineers within the platform organization remained skeptical about the competitive edge of agentic engineering until late 2025, when a new generation of models changed what was possible.

Selecting Claude on Google Cloud’s Vertex AI 

Delivery Hero centralized its approach using Google Cloud's Vertex AI, which hosts Claude and other AI models as a managed service. The company needed a way to give engineers across the entire organization access to the same models without locking them into a single tool, and without months of procurement delays for each new integration. "The tooling itself wasn't the main driver of the success of agentic engineering," Schäfer said. "The main driver was the new models that were substantially better in terms of quality."

Vertex AI solved the access problem. By placing a proxy called LiteLLM in front of the platform, Delivery Hero could route engineers to the best available models regardless of which coding tools they preferred. "Vertex AI allowed us to set up a proxy and give our engineers choice by offering all Vertex AI models, hooked up with whatever tools they use," Schäfer said. "This combination lets us move forward with the newest models. Otherwise, we might need to wait several months."

Once the infrastructure was in place, the preference became clear. A survey of Delivery Hero's CTO group showed nine votes for the Claude setup versus two for the next closest alternative. Usage data confirmed the pattern: across the company's central engineering teams, Claude models account for roughly 95% of all LiteLLM requests.

Herogen: an autonomous agent for software delivery

The biggest product to emerge from this infrastructure is Herogen, an autonomous software delivery agent built on Claude Opus 4.5 as its primary coding model. The origin was scrappy, reflecting Delivery Hero’s approach to intrapreneurship. "We had a strong belief in this product and the technology," Schäfer said. "We pulled in a principal engineer and two staff engineers and put them on this as a virtual team, self-funded within my organization.”

Herogen picks up tasks assigned through Jira, writes the code, runs and iterates on tests, then submits the result as a pull request, a proposed code change for review. A "council of agents" that includes both Claude and Gemini reviews the code from different perspectives before a human does a final check. Using multiple models for review is intentional: it reduces the chance that blind spots in any single model's training data slip through.

The team shipped an MVP with a single merged pull request in Q4 2025. From there, the trajectory was steep. "Our goal for Q1 was 100 merged pull requests," Schäfer said. "We met that before we went GA in February. Then we set 500, met that, doubled to 1,000, met that. We ended the quarter at 18 times the original goal."

Herogen now processes more than 100 merged pull requests per day, a volume that Schäfer noted is what a typical engineer produces in an entire year. Only about 15% of the company's engineers are using it so far. The agent is particularly effective at tasks that span large numbers of code repositories: infrastructure configuration updates and refactoring work that would otherwise require an engineer to open each repo individually. Schäfer's team has seen pull requests that refactored over 50 files in a single repository.

"Agentic engineering replaces rigid programming syntax with the fluidity of natural language,” Schäfer explained. “By embedding this technology into the heart of our software delivery via Herogen, we have moved the primary interface to the change request itself. This allows our engineers to lead with intent rather than code, liberating them to focus on high-level innovation and complex problem-solving. It's about achieving quality as high as possible and going toward zero interactions with developers.”

Claude Code fills a different role alongside Herogen. Where Herogen is designed for autonomous task completion on existing code, Claude Code supports engineers in more exploratory work: building new projects from scratch, experimenting with design approaches, and features that benefit from back-and-forth between the developer and the model. Schäfer described the relationship as one of "symbiosis between the engineer and the agent." One team used Claude Code to overhaul a design system across 100 UI components, a task that would have taken an engineer at least a week. Claude Code finished it in minutes.

Looking ahead

Schäfer is set to expand Delivery Hero’s Herogen team and raise its year-end ambitions significantly. The roadmap includes coordinating changes across multiple connected services, Slack integration for human handoffs, and the ability for engineers to call Herogen into existing pull requests. Beyond Herogen, Schäfer is building a dedicated team that will provide a Claude-based tool stack tailored to Delivery Hero's internal infrastructure, with built-in context for connecting to the company's payment APIs, order APIs, and proprietary infrastructure definitions. The goal is to deploy Claude to approximately 50% of its over 4,000 engineers.

Schäfer is also candid about what comes next. As agentic engineering accelerates the development phase, the constraint shifts downstream. "If you churn out five times more changes to your production environment, you will have five times more incidents if the change failure rate remains at the same level," he noted. The team is investing in agent-driven operations and safer rollouts alongside expanding what Herogen can build.

"Every fifth pull request that's merged, we want that to be done by Herogen in 2026," Schäfer said. "It's a very ambitious target, but we're betting on the improvement in LLMs, especially in Claude, to help us with that."