Pods Work for Execution, Not Deployment
The Pod, a standard in Kubernetes for running workloads, is under closer review for its suitability for AI agents. In a CNCF blog post, Lin Sun argues that while Pods are effective execution environments, they are not ideal for managing how agents are deployed, controlled, and scaled over time.
A key challenge is ensuring each agent operates independently. Each should have its own identity, adhere to network policies, and be separated from other agents. When multiple agents run within the same Pod, tracking and managing them becomes difficult.
The kagent project addressed this by allocating a Pod to each agent. This allows each agent to have a unique Service, ServiceAccount, and dedicated logs and metrics. This improves monitoring and policy enforcement but adds new challenges, including higher resource usage and more complex management.
Pods vs.
The issue is that agents do not run continuously. They may remain idle for long periods and only perform a task for seconds or minutes. Allocating a Pod to each agent results in many idle containers, which waste resources.
Agents may also create subagents to handle parallel tasks or pause for user input. These behaviors do not fit the traditional Pod lifecycle, which assumes steady activity and predictable resource needs.
To address this, Google introduced Agent Substrate, which adds a control plane above Kubernetes. It shifts the management of identity, scheduling, and lifecycle from Pods to a new abstraction known as the Actor.
An Actor is a logical unit that runs on a Worker. Each Worker is linked to one Pod. Unlike a Pod, an Actor can be paused, resumed, or moved between Workers based on needs. This allows a small number of long-running Pods to support many more agents.
This method allows agents to share a Pod while maintaining control. It also moves identity and policy to higher levels, such as ActorTemplates and namespaces. While this may simplify access control, it makes tracking usage, billing, and observability more complicated.
The impact goes beyond scheduling efficiency. Access control, network policies, and runtime permissions may need to be set at the template level, with exceptions for specific Actors. Ownership, quotas, and billing become harder to manage when execution is no longer directly tied to Pods. Observability must now track the logical agent, linking logs, traces, and audit records to the Actor, regardless of where it is scheduled.
Kubernetes remains the standard for managing microservices and inference at scale. Agent Substrate is exploring this through the kagent project. Sun's post was featured in The Kubernetes Podcast's weekly news update.
While Pods remain the foundation of execution, new abstractions like Workers and Actors better match the dynamic nature of AI agents.
The challenge is not just technical. It involves rethinking how workloads are represented and managed, and who is responsible for them. With AI agents acting on behalf of users or performing unpredictable tasks, the need for flexibility and isolation increases.
The goal is to provide efficient and effective execution environments for AI agents while maintaining the robustness, security, and manageability that Kubernetes has long offered. Whether this means moving away from Pods as the primary unit for deployment and lifecycle remains an open question. Actor-based systems like those in Agent Substrate offer a promising direction. As new workloads emerge, the tools and platforms that support them must evolve.
For now, Kubernetes continues to provide foundational infrastructure for deployment, networking, and compute, while layers like Agent Substrate handle the dynamic aspects of managing AI actors. The collaboration between these layers is a step forward in addressing AI workload challenges.
The integration of kagent with Agent Substrate shows how these abstractions can be applied in practice. By using familiar concepts like NodePools and WorkerPools, the transition to these new models can be more intuitive and less disruptive.
As the kagent project continues to evolve, it may offer more insights into balancing Kubernetes strengths with the specific needs of AI agents. The outcome could shape the future of workload management in cloud-native environments.
The broader implications of this shift extend beyond technical efficiency. They involve how responsibility for workloads is defined and managed. With AI actors performing complex and unpredictable tasks, clear ownership and accountability become even more important.
The debate over whether Pods should remain the unit of deployment and identity for AI agents is a key part of the cloud-native computing conversation. As tools like Agent Substrate and kagent develop, they may help set a new standard for managing AI workloads. Kubernetes' role as the industry's leading platform remains intact, but how it supports evolving workloads may continue to change.