The agent workload is structurally completely different
Brokers are long-running, stateful processes. They cause throughout time, name exterior instruments, spawn subprocesses, write and execute code, and make choices that rely on what occurred 5 steps earlier in the identical activity. A single-agent workflow may run for minutes or hours, touching a dozen exterior programs and producing intermediate outputs that subsequent steps rely on. The compute layer for that type of work must do issues the previous mannequin was by no means requested to do. That’s the new sample: Execution infrastructure designed round agent semantics moderately than request semantics.
The Kubernetes neighborhood itself has acknowledged this mismatch. In March 2026, Kubernetes SIG Apps revealed an introduction to Agent Sandbox, a brand new CRD-based abstraction designed particularly for singleton, stateful agent workloads. The framing is direct: The ecosystem is shifting from short-lived, remoted duties to deploying a number of, coordinated AI brokers that run constantly, and mapping these workloads to conventional Kubernetes primitives requires a wholly new abstraction. The truth that the Kubernetes maintainers constructed a devoted primitive for this, moderately than recommending groups compose one from current sources, is itself the clearest sign that agent execution doesn’t match the previous mannequin.
What agent execution really requires
Concretely, it requires 4 issues. First, remoted execution environments that provision in milliseconds, not minutes, so every agent activity will get its personal sandbox for code execution and gear calls with out blocking the reasoning loop. The distinction between a two-second setting and a two-minute setting is just not a efficiency optimization; it determines whether or not the structure is viable in any respect. Second, sturdy state administration throughout the complete activity lifecycle, so an agent can pause, hand off or resume with out re-initializing from scratch and burning tokens to reconstruct context it already constructed. Third, coordination primitives for multi-agent work: The power to spawn subagents, cross structured outputs between them and observe activity dependencies throughout a graph of concurrent processes. Manufacturing agent programs are not often single brokers; they’re pipelines of specialised brokers with handoffs that have to be dependable and inspectable. Fourth, credentials and secrets and techniques administration that journey with the execution context, so brokers can authenticate to exterior providers securely with out exposing credentials within the activity definition, logs or the setting variables of a shared container.