Every time an AI coding agent starts a fresh environment to work in, it waits. Until this week, spinning up a container on Cloudflare took just over four seconds from request to first command. Four seconds. For an agent working on a task while a user watches, expecting near-instant feedback, that is not a minor delay.
That number is now 648 milliseconds. Cloudflare rebuilt the scheduling layer behind its Containers service and cut median startup time by a factor of six, measured in an independent benchmark where 100 containers launched concurrently. The p95 latency dropped from 5.8 seconds to under a second.
The redesign reflects a real difference between how traditional containers get deployed and how AI agents actually work. An agent workspace is different from a standard application. It gets created on demand, often while a user is already waiting, and the task determines the image, resources, and environment it needs. Those decisions cannot be locked in at deployment time.
The new architecture moves those decisions into application code. Previously, choosing a different container image or compute size for different tasks meant creating an entirely separate application deployment, with its own configuration, its own namespace, and routing logic to send tasks to the right environment. It was messy.
Now it is an if statement. The code picks the image and compute size at the moment the task arrives, and a single deployment can spin up Python, Node.js, or any other environment side by side without additional deployment work. Rollout strategies, previously a whole operation involving drain periods and traffic splits, become a few lines of logic in the application itself.
The other significant addition is filesystem snapshots, now in public beta. Snapshotting sounds incremental until you understand what it replaces. Setting up a development environment from scratch, cloning a repository, installing dependencies, and configuring tooling can easily take longer than the container startup itself. Snapshots fix that. A workspace saved at any point restores in milliseconds instead of repeating the full setup process.
This matters for AI evaluation pipelines, which are already a common workload. Running the same scenario across different model versions or agent configurations requires a consistent starting state, and snapshots provide a shared checkpoint that multiple containers can restore from simultaneously, each making independent changes from the same baseline without environment drift.
Real engineering teams are already using this. Kilo Code, which builds cloud-based coding agent sessions, needs isolated workspaces with the right repository, tools, and configuration for each session. They need to be ready immediately.
DevOps engineering is already moving this direction. The teams doing it well are thinking about container infrastructure in terms of per-task startup latency, workspace continuity, and per-task cost, not just uptime and availability. Infrastructure designed for long-running services has always made AI agents awkward to deploy. Infrastructure designed around agent workflows makes all of this the default behavior.
If your team is starting to run AI agent workloads in production, the gap between infrastructure designed for traditional apps and infrastructure built for agents is real, and it compounds fast.
Want to explore how modern DevOps practices could help your development team move faster? Let’s talk.