Cloudflare Dynamic Workflows Brings Per-Tenant Durable Execution to Every Platform
Cloudflare just made a move that platform builders, AI developers, and CI/CD teams will want to notice. On May 1, 2026, the company launched Cloudflare Dynamic Workflows, a small but powerful open-source library. It bridges durable execution with dynamic deployment.
Simply put, it lets any platform run a different workflow for every tenant, without knowing that code ahead of time.
For years, Cloudflare Workflows has offered durable execution. It keeps long-running processes alive through crashes, retries, sleeps, and restarts. However, one hard assumption always existed: workflow code had to ship as part of your own deployment. That meant one class, one binding, one deploy. This model works fine for traditional apps. It breaks down the moment a platform wants customers to bring their own logic.
That’s exactly the gap Cloudflare Dynamic Workflows now closes.
The new library, roughly 300 lines of TypeScript, lets a single Worker act as a loader. It routes every workflow creation call to the right tenant’s code at runtime. Crucially, it also ensures the Workflows engine routes back to that same tenant code when it wakes up, whether seconds, hours, or days later. Workflow IDs, status checks, pausing, retries, and hibernation all work exactly as before. Now, though, each tenant gets their own sandboxed execution environment.
This launch follows a series of related releases. Earlier this year, Cloudflare shipped Dynamic Workers in open beta, a compute primitive that spins up isolated, sandboxed Workers in milliseconds. Then came Durable Object Facets, which extended dynamic storage. Each tenant got their own SQLite database, spun up on demand. After that, Artifacts brought a Git-native versioned filesystem, scalable to tens of millions of instances. Together, these form a stack where compute, storage, source control, and durable execution are all dynamic and tenant-aware.
Cloudflare Dynamic Workflows sits at the top of that stack. It wraps the existing Workflow binding so tenants see a familiar interface. Internally, it tags each workflow with tenant metadata. That metadata routes execution back to the correct dynamic Worker every time the engine resumes a run. The tenant never knows any of this is happening. Their code behaves like a perfectly normal workflow.
The use cases are broad. AI coding agents that write durable execution plans, such as those built on Claude Code or the Cloudflare Agents SDK, can now run those plans as first-class Cloudflare Workflows. Each step is independently retryable. Each sleep hibernates for free. Each human approval gate waits indefinitely without holding a machine open.
For CI/CD platforms, the implications go further. Every repo defines its own pipeline. Every pull request spawns a workflow instance that must survive crashes, retry flaky steps, and wait for deployment approvals. That fits perfectly with durable execution. Until now, no cloud primitive let each repo ship its own workflow at runtime, at zero provisioning cost. Cloudflare Dynamic Workflows makes that possible. The company’s own example code demonstrates a full lint-test-build-deploy pipeline written entirely inside the customer’s repo.
The performance gap compared to traditional CI is significant. A standard run for a mid-sized JavaScript project typically needs 15 to 30 seconds to allocate a VM, 10 seconds to pull a base image, another 10 for a git clone, and up to 60 seconds for dependency installation, all before a single test runs. With Cloudflare Dynamic Workflows, Artifacts forks the repo in seconds. Each step then boots a sandboxed isolate in milliseconds. Between steps, the workflow hibernates rather than holding resources open.
The economics shift just as dramatically at scale. Multi-tenant platforms used to mean giving every customer their own container, database, disk, and scheduler. With Cloudflare Dynamic Workflows and the underlying Dynamic Workers primitive, idle tenants cost effectively nothing. Active tenants share hardware through isolate-level multi-tenancy. As a result, a platform that once topped out at thousands of customers can realistically serve tens of millions.
The library is MIT-licensed and available on npm today. It runs on top of Dynamic Workers, which requires the Workers Paid plan. The GitHub repository includes a working browser-based playground. Developers can write a tenant workflow, run it live, and watch each step execute with streaming logs.
Cloudflare has also signaled that Dynamic Workflows is just the beginning. The company says every binding Workers currently supports, queues, caches, databases, AI bindings, and MCP servers, is heading for a dynamic counterpart. Whatever you bind to a Worker today, you will eventually dispatch per tenant, per agent, or per request, at zero idle cost.
