Sally Ann O’Malley, a Principal Software Engineer at Red Hat, presented a containerization methodology for OpenClaw agents at AI Engineer Europe that addresses one of the core friction points in enterprise AI agent adoption: setups that work on a developer’s laptop but fail in production.
O’Malley’s talk, titled “Lobster Trap: OpenClaw in Containers from Local to K8s and Back,” demonstrated how packaging OpenClaw agents and their full dependency graphs into containers transforms ad-hoc configurations into reproducible, portable assets that run identically across local machines, virtual machines, and Kubernetes clusters, according to StartupHub.
The Reproducibility Problem
The core issue O’Malley addressed is environmental drift. An OpenClaw agent configured on one developer’s machine may depend on specific system libraries, API keys injected through environment variables, model configurations, and skill sets that exist only in that developer’s local environment. Moving that agent to a staging server, a colleague’s machine, or a production Kubernetes cluster often requires rebuilding the setup from scratch.
Containerization solves this by capturing the entire runtime environment, including the agent, its dependencies, secrets references, and configuration, into a single portable image. The same container image runs identically regardless of the underlying infrastructure.
How the Lobster Trap Works
O’Malley’s methodology addresses five specific enterprise requirements, per StartupHub:
Reproducible environments. The same container image ensures all agents operate in identical conditions across development, staging, and production.
Secrets isolation. OpenClaw uses a SecretRef abstraction that supports different injection mechanisms depending on the environment. For local development with Podman, secrets go through environment variables. In Kubernetes, they route through Kubernetes Secrets. API keys for inference providers (OpenRouter, Anthropic, Google) never touch the host system.
Portability. Containers move between environments with minimal changes. A developer testing locally with Podman can push the same image to a Kubernetes cluster without reconfiguration.
Volume-backed persistence. Runtime state and agent data persist on volumes, surviving container restarts or migrations.
Security boundaries. Containers isolate agent execution from the host system, preventing interference with other workloads.
Team Adoption Angle
O’Malley framed the methodology as a team standardization tool, not just an infrastructure convenience. Containerized baselines let new engineers onboard without debugging environment-specific setup issues. Team standards for skills, model choices, and configuration become shared artifacts rather than tribal knowledge, according to StartupHub.
Infrastructure Maturation Signal
The presentation fits a broader pattern in the agent ecosystem’s evolution from “what can agents do” to “how do you run agents reliably.” In the same week, CopilotKit released AIMock for fixture-driven agent testing, and the open-source community began filtering maintained agent frameworks from abandoned repositories. Red Hat’s involvement signals that established infrastructure vendors see AI agents as a workload category worth supporting with production-grade tooling, not an experiment to observe from the sidelines.