Zalando shares a 2.5-year retrospective on scaling agentic engineering across 250+ teams. Topics include an internal LiteLLM-based API proxy with cost tracking and prompt caching, a custom CLI and chat UI built on pydantic-ai, vendor independence via tools like opencode and pi, measurable impact of AI coding on PR sizes and code complexity in Java/Go codebases, a risk-based PR auto-approval bot that cut lead time 20-40%, agent skill libraries, governance through an internal Tech Radar, and knowledge-sharing formats like an LLM guild, hackathons, and GenAI Labs training sessions. The piece closes with future plans including an agent platform built on kagent and an Identity Broker for auth in agentic systems.
Table of contents
LLM proxy for API-based LLM access from day 1Vendor independenceIdentifying the impact of AI codingRisk-based PR approvalLearning from session dataAgent skillsGovernanceKnowledge sharingGetting to the next levelWhat's next?Questions this post answers
How much did a risk-based PR auto-approval bot reduce pull request lead time?
A risk-based PR approval bot that evaluates each pull request as low, medium, or high rollout risk reduced PR lead time by 20-40% compared with all PRs. About 33% of pull requests are classified low-risk and auto-approved, letting authors merge without waiting for a human reviewer, especially useful for prototypes and internal tooling changes. See how daily.dev surfaces engineering practices like automated PR risk scoring for teams speeding up delivery.
What is LiteLLM used for as an API proxy for accessing multiple LLM providers?
LiteLLM is used to run a self-hosted proxy giving unified API access to models from OpenAI, AWS Bedrock, and Google Vertex, letting engineers experiment across providers through one endpoint. Post-call hooks enable anonymized cost tracking, pre-call hooks enforce client version upgrades via User-Agent restrictions, and auto-injected prompt caching checkpoints reduce costs; restarting after 20k requests mitigates memory leaks. Developers comparing LLM gateway tools can track proxy and infrastructure patterns like this on daily.dev.
Does adopting AI coding agents increase code complexity and pull request size?
Yes, analysis of commit-level cyclomatic complexity across Java and Go codebases shows clear inflection points in complexity growth coinciding with coding agent adoption, alongside consistent increases in PR sizes, especially larger buckets like 500-1k and 1k-2k lines since the Sonnet 4 model release in Q2 2025. Codebases built agentic-only from day one show complexity building up quickly before growth fades. Track real-world evidence on how coding agents affect codebase health by following engineering stories on daily.dev.