Delivery Hero's Agentic Engineering Strategy
Delivery Hero's SVP of Tech Foundations details the shift from decentralized AI experimentation to a centralized agentic stack. The company achieved 18x its initial Q1 targets for merged PRs by integrating autonomous agents into CI/CD pipelines. This analysis covers governance, cost control, and the roadmap to a fully autonomous software factory.
The Shift to Centralized Agentic Stacks
Delivery Hero’s transition from decentralized AI experimentation to a centralized agentic stack represents a critical maturity milestone for enterprise software engineering. Initially, various subsidiaries experimented with disparate tools like Cursor, Windsurf, and Claude, leading to fragmented governance, unclear ROI, and security vulnerabilities. By consolidating around a single vendor (primarily Claude) for central teams, the company established the necessary infrastructure for compliance, cost control, and measurable impact. This strategic pivot highlights that while decentralized innovation is valuable for discovery, centralized governance is essential for scaling AI tools across thousands of engineers.
Quantifiable Impact on Engineering Velocity
The implementation of the "HeroGen" autonomous agent has yielded extraordinary results. In Q1, the system generated 1,800 merged pull requests against an initial target of 100, an 18x overachievement. More importantly, the quality of this output is high, with an 83% merge rate in September, indicating that the majority of agent-generated code is production-ready without significant human modification. This data challenges the prevailing skepticism regarding AI code quality, suggesting that with proper harnessing and context, agents can deliver reliable, high-volume software changes.
Operationalizing the Software Factory
The next phase of this transformation involves moving from assisted coding to an end-to-end "software factory." This requires solving two major operational challenges: review bottlenecks and incident management. As agent-generated PRs increase, human review capacity becomes a constraint. Delivery Hero is addressing this by implementing risk-based auto-approval for low-risk changes, supported by rigorous DevOps safety nets like canary rollouts and automated rollbacks. Furthermore, the increased deployment frequency necessitates autonomous incident handling. If human engineers remain the sole responders to production issues, the system’s resilience will fail. The company predicts that by the end of next year, agents will handle a significant percentage of operational incidents, ensuring that the speed of code production does not outpace the organization’s ability to maintain stability.
Strategic Implications for Leaders
For CTOs and engineering leaders, this case study underscores the importance of building "context hubs" to provide agents with architectural visibility. Without understanding cross-service dependencies, agents cannot make safe decisions. Additionally, the cost management of AI usage requires granular tracking and limits, moving beyond simple subscription models to usage-based controls. The trajectory is clear: the competitive advantage will shift from who has the best model to who has the best operational harness for autonomous software delivery.
Key insights
-
Decentralized AI adoption creates governance and security risks that scale poorly. Centralizing the tooling stack is necessary to enforce compliance, track ROI, and manage costs effectively in large organizations.
Impact: Reduces security vulnerabilities and provides clear data on AI investment returns, enabling better budget allocation and risk management.
-
Autonomous agents can significantly outperform human-only workflows in volume and speed. Delivery Hero’s 18x overachievement of PR targets demonstrates the massive throughput potential of agentic systems.
Impact: Accelerates feature delivery and reduces time-to-market, allowing companies to iterate on products faster than competitors relying on traditional manual coding.
-
High merge rates (83%) for agent-generated code indicate that modern LLMs, when properly harnessed, produce production-quality code. This shifts the bottleneck from code generation to code review and deployment.
Impact: Increases confidence in AI-generated code, enabling organizations to automate more of the development lifecycle and reduce reliance on manual coding tasks.
-
Scaling autonomous deployment requires risk-based auto-approval mechanisms. Without these, human review becomes a bottleneck that negates the speed benefits of agentic coding.
Impact: Enables true end-to-end automation of the software delivery pipeline, reducing cycle times and allowing for more frequent, smaller deployments.
-
Agents lack inherent visibility into complex microservice architectures. Building centralized context hubs that map dependencies is critical for agents to make safe, informed architectural decisions.
Impact: Improves the accuracy and safety of agent-generated code by providing necessary architectural context, reducing the likelihood of breaking changes or integration errors.
Action items
-
Audit current AI tool usage across engineering teams to identify fragmentation. Consolidate tools into a centralized stack to improve governance, security, and cost visibility.
Impact: Enhances security posture and provides a single source of truth for AI usage metrics, facilitating better strategic decision-making.
-
Implement risk-based classification for pull requests. Define criteria for low-risk changes that can be auto-approved, supported by robust automated testing and rollback mechanisms.
Impact: Reduces review bottlenecks and accelerates deployment cycles, allowing the organization to fully leverage the speed of agentic code generation.
-
Develop a centralized context hub that maps service dependencies and architectural patterns. Integrate this context into agent prompts to improve decision-making accuracy.
Impact: Increases the reliability of agent-generated code by ensuring agents have a comprehensive understanding of the system architecture, reducing integration errors.
-
Establish granular cost controls and usage limits for AI tools. Move beyond flat subscriptions to usage-based tracking to monitor ROI and prevent cost overruns.
Impact: Optimizes AI spending and ensures that investment in agentic tools delivers measurable value, preventing budget blowouts due to uncontrolled usage.
-
Invest in autonomous incident handling capabilities. Develop agents that can monitor production environments and automatically resolve common operational issues.
Impact: Prevents human engineers from becoming a resilience bottleneck as deployment frequency increases, ensuring system stability at scale.
Quotes
“In der Ende, wir haben 18x den Original Target in Q1, weil es wirklich skyrocketet in terms of Usage.”
“So we wanted to centralize that. We want to have some sort of technological ownership. On that topic.”
“And a bottleneck that's not about just capacity, but it's really about resilience.”