From Containers to Sessions: The New AI Agent Runtime Layer
An AI agent runtime is an infrastructure layer that manages long‑running, stateful agent sessions, giving each autonomous worker persistent memory, secure code execution, and event‑driven scheduling beyond traditional container orchestration. Instead of treating agents as short‑lived services, these runtimes treat them as sleeping processes that wake on demand, preserve context across hours or days, and cooperate with other agents and humans in shared workflows. The key takeaway is blunt: Kubernetes may still be the right tool for provisioning machines, but it is the wrong control plane for persistent AI agents that nap most of the time and burst into activity when a prompt or event lands. The rise of Google’s Agent Substrate, Raft 1.0, and Perplexity SPACE marks a pivot from ephemeral compute to persistent agent orchestration.
Google’s Agent Substrate: Admitting Kubernetes Isn’t Built for Agents
Google made GKE Agent Sandbox generally available and introduced Agent Substrate in the same announcement, a quiet concession that Kubernetes was never designed for how agents behave. Agent Sandbox is an open‑source, sandboxed runtime that gives each agent a hardened environment for model‑generated, untrusted code, treating the workload as hostile and enforcing isolation at the kernel boundary rather than the container boundary. Agent Substrate then routes around the Kubernetes API server, adding a scheduler tuned for millions of mostly idle sessions instead of a modest number of long‑lived Pods. Google reports the API can allocate 300 sandboxes per second per cluster, with 90 percent completing in 200 milliseconds, after roughly 16x growth in GKE sandboxes in under five months. That performance is not a nice-to-have; it is a direct response to workloads where a coding agent may sit open for twenty minutes, then run code for ten seconds when a prompt arrives.
Raft 1.0: Turning Multi-Agent Workflows into a Shared Workplace
Where Agent Substrate tackles cluster‑level scheduling, Raft 1.0 tackles the human side of agent orchestration platforms. The release marks the full launch of a multi‑agent collaboration system built around "team mode," replacing a mess of terminals and agent windows with one shared workspace of channels, threads, tasks, and mentions. Each agent on Raft runs as a persistent process with its own identity, memory, and expertise, on whichever runtime fits the job, including Claude, Codex, and Hermes. Agents claim tasks, run in parallel, hand work to one another, and review each other’s output in shared threads, so multi-agent workflows feel like a team, not a pipeline. A separate review agent can catch issues in another agent’s code before a human sees it, framing independent agents as covering independent blind spots. For ordinary users, Raft is now available with a free tier that includes channels, tasks, local agents, and 30 days of message history, with an enterprise tier with private deployment and SSO listed as coming soon.
Perplexity SPACE: Sandboxed Runtime for Long-Running Agents at Scale
Perplexity’s SPACE runtime is a clear statement that long‑running agents need a sandboxed runtime designed for months‑long work, not web request lifecycles. SPACE, short for Sandboxed Platform for Agentic Code Execution, supports agents that execute code, edit files, and complete multi‑step tasks over hours, days, or even months. Each SPACE sandbox runs inside a virtual machine with its own guest kernel, separating untrusted workloads from the host and other tenants while a dedicated in‑guest daemon controls filesystem and process operations and a gateway governs outbound network traffic. The platform’s storage layer, built on copy‑on‑write snapshots, gives agents persistent AI memory with pausing, resuming, suspension, restoration, rollback, forking, and crash recovery. In production testing, median sandbox creation time dropped from 185 milliseconds to 60 milliseconds, and 90th‑percentile latency fell from 447 milliseconds to 89 milliseconds, making SPACE about 3.1 to 5 times faster than the prior sandbox system. Perplexity is already using SPACE for all Computer sessions and plans to expand it to Linux microVMs, Windows guests, and users’ local machines.
Why Persistent Agent Orchestration Is Replacing Ephemeral Compute
Taken together, Agent Substrate, Raft, and SPACE show the same pattern: traditional container infrastructure is misaligned with what autonomous agents need. Agents are long‑running, stateful sessions that stay idle most of their life, waking to execute a burst of code before going quiet again. They need persistent memory across naps, secure isolation for model‑written code, and coordination across many agents and humans using shared workspaces and multi-agent workflows. These runtimes move orchestration away from ephemeral Pods and Lambdas toward persistent sessions, snapshots, and fine‑grained scheduling tuned to event‑driven agents rather than HTTP traffic. The practical impact is already visible: developers keep coding agents open all afternoon without burning cluster resources, teams collaborate with fleets of agents inside Raft, and Perplexity users run long‑lived agents that can pause, resume, and even fork their work. Kubernetes will remain the data center substrate, but the future of agent orchestration belongs to these purpose‑built AI agent runtimes.
