I am not a Claude pair programmer — I am an agent manager. My role is to enable the agents, protect the environment they run in, and verify their results automatically. Several Claude Code instances run in parallel on my watch, each in its own sandbox on its own branch; some work through the night.

Protection is structural: each instance runs in a isolated execution environment with an firewalled network and a filesystem it cannot read outside its task. Tooling is in-house — 8 skills I authored, on-demand rule packs, a memory MCP for long-running work, and Chrome, Playwright, and iOS Simulator MCPs so tasks can be described in what a user sees rather than which functions to call. Verification is automated: cross-agent deliverable reviews, unit and integration tests in an isolated container, visual diffs, and behaviour reproducible outside the session. Multi-agent pipelines produce intermediate results I can check, so I am never reviewing a single opaque patch at the end.

The outcomes are measurable. Developers ship 2–3× more features, hand off shareable bug reproductions to other teams in under half an hour without pausing their own work, and let agents close simple bugs or run long research and migration tasks overnight. PR volume rises; the verifiable-deliverables pipeline and peer AI reviews — driven by our team's AI rule framework — hold the quality bar. A couple of minor incidents from agent-generated code along the way, none of them sandbox escapes, all caught by monitoring, each one folded back as a stronger safeguard. A fair price for the productivity gain. 3+ years of hands-on experience.

As a tech leader I act as the connective tissue between backend, design, dev-ops, legal, product, data, HR, and PR — the person teams call when the answer spans two codebases or crosses team boundaries. I author the ADRs that outlive individual features: federation schema contracts, deprecation policy, AI-governance policy, supply-chain boundaries, query-complexity governance.

Across senior, principal, and staff roles in Germany I have led 8 teams and shipped in both modes — greenfield (four generations of products at Matterway; a crypto indexer and cross-team Federation Router in my current role) and brownfield, where the hardest call is usually what not to rewrite.

I stay hands-on. The ADRs are load-bearing because I still write the code I sign off on. 19+ years in the industry, 8 teams led.

Mentorship and knowledge-sharing are central to how I operate, not a side activity. I have mentored 114+ engineers across my teams, taught JavaScript and web technologies for two years at Yandex Interface School, and delivered 13 public talks at international conferences.

Inside my current company I created and run the #ai-workbench channel — shared skills, rule packs, review patterns, and latest AI news — reaching almost every developer on staff. I also maintain shared component libraries and coding standards as the teaching artifacts people actually use day to day, rather than as policy documents nobody reads.

My bias is simple: short feedback loops, show rather than tell, and invest in engineers who will outgrow me. Good mentorship is measured by what your mentees ship after you stop reviewing. 114+ engineers mentored, 13 public talks.