A meetup series for platform engineers, DevOps practitioners and the cloud native community, focused on what happens when coding agents move from individual laptops into shared infrastructure.
The second edition tackles a problem every platform team is about to own. Skills, MCP servers and CLAUDE.md files multiply faster than anyone can review them. Shared naively, they turn into an ungoverned pile. Two speakers present two production-tested answers: composable specialist agents governed centrally, and layered, team-owned context treated as code inside a large monorepo.
Stop Sharing Skills: Composable Agents for Your Team
Dominik Schmidle · Product Lead for Platforms, Giant Swarm
Most of us have a growing pile of skills, MCP servers and CLAUDE.md files on our laptops. Sharing them just creates an even bigger pile in the team account. In my experience, the thing worth sharing is not a prompt or a skill, but a specialist agent, composed from building blocks that are easy to provide and govern centrally.
In this talk I’ll show what a platform for managing those building blocks and a whole fleet of agents can look like.
Dominik Schmidle is Product Lead for Platforms at Giant Swarm, where he builds the Agent Platform that lets organisations adopt AI agents in an easy and governable way.
Designing an Onion: Domain-Scoped AI Context at Monorepo Scale
Na’aman Hirschfeld · Principal Systems Engineer, ServiceNow (Armis)
Coding agents are only as good as the context they are given. In a large monorepo, dozens of teams write that context: hundreds of skills, rules and instructions that grow faster than anyone reviews them. How do you keep it correct?
We started with plugin marketplaces, one per team, and quickly hit the limits of the harness. Engineers ended up running different subsets, and no two agents worked from the same rules. So we rebuilt the setup as an onion. A small core loads in every session. Each directory beneath it adds its own layer of rules, skills and instructions. The team that owns the code owns that layer, and it loads only when the agent reaches into that part of the tree.
We treat this as a developer-experience product, because engineers are our users. Shared MCP servers ship with the repository, and a repo-wide skill index lets agents find what they need without reading the whole tree. We also shifted quality left. AI configuration is code: it gets linting, a sandboxed skill scanner, an LLM judge, and dedicated review rubrics applied by our PR reviewer.
This talk covers the layered architecture, the loading behaviour we measured, the review and enforcement chain, and how we migrated in small, reviewable steps.
Na’aman Hirschfeld is a Principal Systems Engineer at ServiceNow (Armis), creator of the open-source libraries Xberg, Litestar and polyfactory, and the CTO of his own startup Xberg.io. LinkedIn · GitHub
