How Delivery Hero's Herogen Agent Merges 100+ Pull Requests a Day with Claude

Delivery Hero operates in about 65 countries, processing over 11 million orders daily through local delivery brands. Since its founding in 2011, it has grown largely through acquiring regional delivery leaders—each of which came with its own cloud providers, programming languages, and engineering culture. When the company centralized its developer platform starting around 2021, it inherited this fragmentation across thousands of engineers.

AI tooling adoption faced the same challenge. Starting in mid-2024, Delivery Hero ran a decentralized experimentation phase, letting teams pilot different solutions. The results were promising but hard to measure consistently. Skepticism about agentic engineering persisted within the platform organization until late 2025, when a new generation of models changed what was possible. "The tooling itself wasn't the main driver of the success of agentic engineering," said Rodrigue Schäfer, VP of Platform. "The main driver was the new models that were substantially better in terms of quality."

The breakthrough came from centralizing model access through Google Cloud Vertex AI with a LiteLLM proxy in front of it. Engineers kept their preferred coding tools; the proxy routed them to the best available models. "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. The preference rapidly became clear: a survey of Delivery Hero's CTO group showed a 9-to-2 vote for the Claude setup, and Claude models account for roughly 95% of all central engineering LiteLLM requests. The biggest product to emerge was Herogen—built by three engineers self-funded within Schäfer's organization. Herogen picks up Jira tickets autonomously, writes the code, runs and iterates on tests, then submits a pull request. A "council of agents" using both Claude and Gemini reviews the code from different perspectives before a human does a final check. The MVP shipped with a single merged PR in Q4 2025.

The ramp was steep. The Q1 2026 goal was 100 merged PRs. Herogen hit that before going GA in February, so the team raised the target to 500, hit that, doubled to 1,000, hit that again, and closed the quarter at 18x the original goal. Herogen now merges more than 100 PRs per day—roughly the annual output of a typical engineer—with an 85% success rate on assigned tickets and most tasks completed with zero to one developer interactions. Engineers can now initiate cross-repository changes across hundreds of codebases—work that previously took multiple days—in minutes. Claude Code fills a complementary role for exploratory work, new projects, and features that benefit from back-and-forth. One team used it to overhaul a design system across 100 UI components in minutes rather than a week.

"Every fifth pull request that's merged, we want that to be done by Herogen in 2026," Schäfer said. Only about 15% of Delivery Hero's 4,000+ engineers are using it today. The next phase includes cross-service coordination, Slack integration for human handoffs, and expanding Herogen to approximately 50% of the engineering organization. "Agentic engineering replaces rigid programming syntax with the fluidity of natural language," Schäfer said. "This allows our engineers to lead with intent rather than code, liberating them to focus on high-level innovation and complex problem-solving."