Cloudflare launches Dynamic Workflows to run tenant-specific durable workflows at runtime

Cloudflare on May 9, 2026 announced Dynamic Workflows, a small MIT‑licensed TypeScript library that extends its durable execution engine so platforms can run durable workflows whose code varies per tenant, agent, or individual request at runtime. The feature lets platforms route execution to tenant — specific modules without changing core orchestration semantics, enabling per-tenant customization while preserving durable, retryable execution — a capability that matters for CI/CD, multi — tenant apps, and agent orchestration at scale.

At the technical level the library inserts a Worker Loader between the Workflows engine and tenant code (a Dynamic Worker). When tenant code invokes env.WORKFLOWS.create(...), the loader wraps the call with tenant metadata; the Workflows engine persists the payload and later rehydrates execution using that metadata to route run(event, step) back to the correct tenant’s module. Existing Workflows semantics — IDs, pause/resume, retries, hibernation, step.sleep('24 hours'), and step.waitForEvent() — remain unchanged. Cloudflare says the implementation is compact, roughly 300 lines of TypeScript.

Cloudflare illustrates several concrete platform patterns that Dynamic Workflows enables. In an app platform scenario, AI could generate TypeScript per tenant and have the Workflows engine orchestrate tenant — specific logic. In an agents SDK each agent can carry its own durable plan. The most prominent example is CI/CD: every repository can hold its pipeline code and use Dynamic Workers to execute steps while the Workflows engine manages durable orchestration across pauses and approvals.

In the CI example the pipeline entrypoint lives inside the customer repository; Dynamic Workers run lightweight steps such as lint, typecheck and bundle inside sandboxed isolates while the orchestration state persists in the Workflows engine. Cloudflare pairs Dynamic Workflows with other primitives to reduce the overheads that traditionally make CI slow: Artifacts provides Git‑native versioned storage with lazy tree hydration and instant fork() per run; Dynamic Workers deliver millisecond isolate boot times; Sandboxes take heavyweight steps with snapshot‑based warm starts.

The company contrasts this approach with the VM/provisioning ceremony — allocating a VM, pulling images and installing dependencies — that can add minutes before tasks start. For platform builders the practical implications are scale and cost. Loading code and repository trees on demand lets operators route millions of unique durable workflows while keeping idle cost low, because tenant code is only materialized when needed. Headline use cases include CI/CD and agent plan execution, and the library’s small size plus its MIT license lower friction for teams that need per‑tenant dynamic code combined with durable, retryable execution semantics.