Cloudflare Rearchitects Workflows for AI Agents

Cloudflare revamps its Workflows control plane with a new architecture to handle the massive scale demanded by AI agents.

3 min read
Diagram illustrating the V2 architecture of Cloudflare Workflows control plane with SousChef and Gatekeeper components.
An overview of the rearchitected Cloudflare Workflows control plane.· Cloudflare

Cloudflare is overhauling its Workflows control plane, a move necessitated by the dramatic shift from human-initiated tasks to the relentless pace of AI agents. Originally designed for user actions like sign-ups, Workflows now must handle persistent, autonomous agents operating at machine speed. This evolution demands a more robust, asynchronous execution engine capable of managing thousands of instances concurrently.

The company is now supporting 50,000 concurrent workflow instances, up from 4,500, and can create 300 instances per second per account, a threefold increase from the prior 100. Millions of queued instances are also now supported. This scaling is achieved through a redesigned control plane, moving from a single Durable Object bottleneck to a horizontally scalable architecture.

The Agentic Shift

The rise of persistent AI agents, which operate on behalf of users for extended periods, requires a durable execution engine. Workflows provides this by ensuring each step is independently retryable, allowing for human-in-the-loop approvals, and surviving failures without losing progress. Developers are increasingly using Workflows to build agent loops and manage their lifecycle.

Cloudflare's Agents SDK integration has further accelerated this trend, allowing agents to easily spawn workflow instances and retrieve real-time progress. A single agent session can now initiate dozens of workflows, and with multiple agents running, thousands of instances can be created in mere seconds.

V1 Limitations and V2 Architecture

The original V1 control plane relied on a single account-level Durable Object to manage all workflows. This proved to be a bottleneck as customer usage scaled, leading to rate limits of 4,500 concurrent slots and 100 instance creations per 10 seconds.

Version 2 of the control plane addresses these limitations with a new architecture built on two key components: SousChef and Gatekeeper. SousChef manages metadata and lifecycle for subsets of instances within a specific workflow, distributing the load. Gatekeeper acts as a leasing system, distributing concurrency slots across SousChefs to prevent overloading the system.

This distributed approach ensures that the source of truth for an instance's existence is its Engine, and lifecycle mechanisms are horizontally scalable. The new Account singleton now stores only minimal metadata, maintaining an invariant maximum number of concurrent requests.

Seamless Migration

Migrating millions of existing instances and thousands of customers to the new V2 control plane without downtime was a significant challenge. Cloudflare achieved this by transitioning existing Account Durable Objects to behave like SousChefs. Instance metadata was persisted, and the Account DO was effectively converted, allowing the new components to be integrated seamlessly into older accounts.

New instance creation requests are now routed to the new SousChefs, while existing Account Durable Objects gradually migrate their state. This process was completed with no disruption, akin to changing a car's wheels while driving, according to the announcement.

Developers can now leverage these enhanced capabilities for their agentic workflows. For those new to Workflows, a Get Started guide is available.

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.