Redesigning the Workflows Control Plane for the Agentic Era

The existing Workflows was designed on the assumption of human-triggered tasks, but now it's an era where agents create and execute workflows at machine speed. A single agent session can spin up dozens of workflows simultaneously, and when multiple agents move in parallel, thousands of instances can be created within seconds.

In response, Cloudflare redesigned the Workflows control plane to scale to the next level.

• 50,000 concurrent instances
• 300 instances/second per account
• 2 million queued instances per workflow

In V1, a single account-level Account Durable Object handled registration, creation, updates, and listing for all workflows and instances. This structure was simple, but became a bottleneck under high concurrency, leading to a hard cap of 4,500 concurrency and 100 creations per second per account.

The core principles of V2 are twofold. First, the existence of an instance should have only the individual Engine as the source of truth. Second, the Account should hold only minimal metadata, and the number of requests it can process at once should always be limitable.

To achieve this, two new components were introduced. SousChef acts as the account's "second-in-command," taking charge of a portion of a specific workflow's instances and distributing metadata and lifecycle management. Gatekeeper is a mechanism that allocates concurrency slots via a lease approach, enabling all SousChefs to efficiently share the account's slots. Slot requests are batched on a 1-second cycle so that calls to the Account don't spike, while also maintaining fairness by prioritizing the resumption of older instances over new ones.

The new creation path is simpler. It checks the control plane version, and if there's no cached workflow information, it fetches and stores the name, unique ID, and version from the Account, then leaves only the necessary metadata on the Engine. Afterward, background jobs and Durable Object alarms guarantee instance startup, keeping the hot-path fast while ensuring reliability through an at-least-once model.

Thanks to this structure, instance listing shows consistent performance that aligns well with cursor pagination, and instance-related operations go directly to the Engine, completing in a single network hop. It also became possible to monitor whether more instances are operating on time simultaneously, and to correct immediately if they fall behind.

Migration was another challenge. With millions of existing instances and thousands of customers already in place, a zero-downtime transition was required. In the end, the existing AccountOld was made to behave like a SousChef, absorbing it into the new architecture while preserving state—and in the process, also resolving an issue from V1's technical debt where some queued instances still had no Engine.

As a result, Workflows shifted from a human-centric, limited-concurrency model into a distributed control plane capable of handling tasks created in bulk by agents and automation. Cloudflare's core message is clear: workflows in the agentic era can no longer just "scale moderately"—they must be systems that safely coordinate hundreds of thousands of states simultaneously.