This post is an organization-scale retrospective rather than an argument about what agentic engineering ought to be. Written by Bartosz Ocytko, an Executive Principal Engineer at Zalando, it looks back over what the company describes as two and a half years of work across more than 250 engineering teams and sets out the approaches it says worked. Its framing is that the industry as a whole is still figuring out how to approach agentic engineering, and that value and impact from large language models are arriving in different forms and at different paces across the company’s business lines.
What distinguishes the account from vendor material on the same subject is that most of its claims are about Zalando’s own measurements and its own operational choices: how many pods the LLM proxy runs on, what fraction of pull requests an approval bot auto-approves, what happened to code complexity in four named codebases once coding agents entered them. The post is explicit that it is a snapshot of an evolving practice and ends by inviting other non-vendor engineering teams working on similar problems to get in touch.
The organizing decision it describes is a deliberate refusal to converge. With more than 200 teams exploring the ecosystem, the post states the company’s position that it is far too early to standardize on one tool, and that its objective instead is transparency and exchange between teams.
Key Points
- Zalando’s ML platform team deployed a LiteLLM-based API proxy in January 2024 to give engineers API-based access to models from multiple providers, giving the platform team a single point at which to measure adoption.
- The company has never centrally mandated a single coding tool; users choose based on available models and their own preference for IDE or CLI, and the post frames vendor independence as key in a fast-moving environment.
- Users become attached to the coding agent they have been using despite what the post describes as rather low switching costs, since the tools’ capabilities are largely similar — the author presents this hesitance as psychological rather than technical.
- Zalando reports seeing the impact of AI coding in its pull-request data for two years: a consistent increase in PRs of 100–500 changed lines, plus growth in the 500–1,000 and 1,000–2,000 buckets since the Sonnet 4 release in Q2 2025.
- Mapping commit-level code quality metrics across four Java and Go codebases, the company says it can pinpoint inflection points in total cyclomatic complexity at the time coding agents entered each codebase; in codebases that used agents fully from the start, complexity builds up very quickly and then fades out. This is the company’s own analysis of its own repositories.
- Commit messages themselves carry a footprint of coding agents, typically around the 5,000 character mark — in one case a commit message included a full unit test execution log.
- A risk-based PR approval bot classifies each PR as low, medium or high rollout risk at creation time; the post reports that 33% of PRs are low-risk and auto-approved, which it says reduced PR lead time by 20–40% compared with all PRs. Its rule set was built from analysis of the company’s own production incidents and is described as highly specific to its tech stack.
- The post reports anecdotal evidence that the approval bot changed engineers’ behaviour: PRs began to be split so that backwards-compatible changes ship quickly as low risk, where previously such changes were mixed with riskier ones.
- Zalando maintains a centralized collection of Agent Skills grouped into plugins, addressing common tasks across disciplines and languages; the post says migration skills — guiding teams through adopting new platform tools or infrastructure practices — are a widely popular type.
- Session data from coding agents is described as highly educational, both for spotting non-essential traffic that costs tokens (generating plan names, terminal window titles, recaps for idle sessions) and for letting users learn about their own prompting patterns.
- Governance is handled through existing mechanisms rather than new ones: an AI section was added to the company’s Tech Radar, and AI model usage is auto-detected by scanning deployed Docker images, after which owners are asked for documentation or a legal review.
- On training, the post states that participants are strongly tempted to use coding agents as a shortcut in sessions meant to build skills, and that doing so usually inhibits learning — so trainers are advised to state explicitly when manual coding is expected.
Context
The post positions itself against accounts from vendors and from monorepo-heavy organizations. It notes that many reported AI wins and increases in PR throughput across the industry come from monorepos where leverage is high, whereas Zalando largely uses separate repositories for its microservices — which is the stated reason it plans a scanner to assess each repository’s AI readiness.
Its caveats are mostly about the limits of its own evidence. The complexity analysis relies on
commits carrying Co-authored-by markers to confirm inflection points, and the author notes that
open-source codebases are less consistent in this because not all authors disclose coding agent
use. On the risk-based approval bot’s effect on behaviour, the post says explicitly that the
evidence is anecdotal. The prediction that complexity plateauing for a well-scoped microservice
means build time has been drastically reduced is stated as a hope, with the author writing that
time will show whether this is the case.
The closing theme is that AI amplifies both good and bad existing practice rather than changing it in one direction: teams that get carried away end up with large PRs that discourage reviewers and slow delivery until the team adjusts, while investments in platform capabilities — such as the web monorepo’s per-PR deployment wired to live data — pay off by letting non-engineers prompt changes and review results before handing over to engineers.