Zalando's 2.5 years of agentic engineering across 250+ teams: a risk-based bot auto-approves 33% of PRs and cuts their lead time 20-40%
Zalando's executive principal engineer describes 2.5 years of agentic engineering across more than 250 teams. Engineers pick their own tools (GitHub Copilot, Claude Code, OpenCode, pi) through a LiteLLM proxy serving about 2,000 monthly users on six small pods. PR data shows PRs growing larger since Sonnet 4, and commit-level complexity metrics show inflection points when agents arrive. A risk-based approval bot, with rules drawn from past production incidents, rates each PR low, medium or high risk and auto-approves low-risk ones. A central skill collection carries migration skills, and a fleet tool now runs coding-agent transformations across many repositories.
The numbers (as reported)
| Teams | More than 250 engineering teams |
|---|---|
| Auto-approved PRs | 33% rated low risk and auto-approved |
| PR lead time | Reduced 20-40% for auto-approved PRs |
| LLM proxy | About 2k MAU on six small pods |
Facts
- Company / subject
- Zalando
- Tools
- GitHub Copilot, Claude Code, OpenCode
- Codebase
- Not applicable
- Work type
- Company rollout
- Stack
- Java, Go, LiteLLM, pydantic-ai, Kubernetes
- Published
- 2026-08-14
- Last checked
- 2026-11-09
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- A risk-based bot auto-approves 33% of PRs 33%
“33% of our PRs are low-risk and are auto-approved by the bot”
- Auto-approval reduced PR lead time by 20-40% 20-40%
“reduced PR lead time by 20-40% (when compared with all PRs)”
Caveats
Written by Zalando's own engineering leadership. The 20-40% lead-time cut is for auto-approved PRs compared with all PRs. No productivity or cost totals; complexity findings come from four example codebases. Also mentions pi, a coding agent not in the site's tool list.
More stories by tool: Claude CodeGitHub CopilotOpenCode
Related topics: Company rollouts and cost