Zalando auto-approves 33% of PRs via risk rules, cutting lead time 20-40%

bibryam · x · 2026-10-29

Zalando's engineering team published a retrospective on 2.5 years of agentic engineering:

Rule-based PR auto-approval: internal risk classification rules derived from their own incident history auto-approve low-risk PRs — 33% of all PRs — cutting lead time by 20-40% versus the overall baseline.
LLM proxy from day one: a LiteLLM-based API proxy deployed in Jan 2024 unifies access to OpenAI, AWS Bedrock and Google Vertex, letting 250+ engineering teams experiment while the platform tracks adoption via MAU/WAU/model/User-Agent.
LiteLLM extension tricks: post-call hooks for anonymized cost tracking; pre-call hooks to enforce client upgrades by restricting proxy access based on the User-Agent header — blocking access is the only effective measure against long-tail users who never update configs, and likewise for retiring models.

More from coding & agent