Grab's CyberSecurity team built Palana, a Kubernetes-native platform for running autonomous AI agents safely at scale. As AI agents evolved from IDE plugins into long-running workloads with network access, credentials, and persistent state, the traditional model of running them on developer laptops became inadequate. Palana addresses this by giving each agent an isolated Kubernetes namespace with scoped RBAC, storage, and network policies. Key design principles include: isolation as the unit of trust (each agent gets its own namespace and Vault scope), proxy-mediated credential injection (agents never see real credentials, only placeholders replaced at the proxy layer), observable egress via Envoy and OPA policy checks, and a control plane that operates entirely outside the agent process. The platform currently runs hundreds of agents including remote dev environments, Slack automations, and long-running task agents. Part 2 will cover the internal architecture in more detail.
