Mick Jermsurawong

Coding agents can rewrite an ML pipeline in minutes. The harder problem is running those changes against real data and distributed compute without slow CI/CD for every experiment—or giving an unattended agent broader access than its human owner. We show how we built an agent-ready Flyte platform on Kubernetes that overlays changed source files across a workflow’s Python-project dependency graph. Engineers and agents can modify workflow orchestration, shared libraries, and model components together, then run the full pipeline remotely without rebuilding an image after every eligible change. Lifecycle-aware skills guide agents through workflow creation, execution, failure diagnosis, and experiment comparison. A read-only MCP interface provides execution context, while persistent development environments enable unattended experiment loops under the requesting human’s identity. We then explain the security model that allows this workflow to scale across enterprise teams: trusted human-to-workload identity propagation, team-scoped permissions, separate production execution identities, and onboarding through existing directory groups. We close with the hardening that makes these boundaries hold—approved images, scoped secrets, workload isolation, and reviewed production promotion. In the latest four-week period, successful top-level workflow executions reached more than 24,000—nearly triple the preceding four weeks. Attendees will leave with practical patterns for fast agentic experimentation without surrendering enterprise control.