Workflow Infrastructure

Automation systems you own — not workflow tools you rent.

The platforms underneath real teams' automation: n8n, Make, Zapier, Pipedream, Activepieces, Windmill. Treated as infrastructure, not a catalog — with the trade-offs that actually matter at scale: ownership, operational cost, migration risk, and the workflow ceiling.

Workflow complexity — where each system tops out

Real automation systems aren't measured in “triggers connected”. They're measured in branches, retries, sub-workflows, memory, and how gracefully the platform fails when something does.

L1

Linear glue: trigger → step → action

One-shot integrations between SaaS apps. Every platform handles this; the question is what they cost when there are 50 of them.
L2

Branching, retries, sub-workflows

The first ceiling for legacy task-priced tools. n8n, Pipedream, Make, Windmill, Activepieces survive here; Zapier starts costing real money.
L3

AI-augmented workflows with memory and retrieval

Embeddings, vector stores, RAG, classification. Built into n8n; possible with code on Pipedream/Windmill; bolted on (and expensive) elsewhere.
L4

Operational automation systems

Production workloads with error workflows, queue mode, cost ceilings, observability, and access control. n8n self-hosted or code-first platforms only.

By system role

Pick by what each platform is actually for, not by feature checklist.

Default substrate

n8n. Visual + code, self-hostable, AI-native, JSON-portable. The recommended starting point for new automation systems.

Code-first cloud

Pipedream. When your team writes JS/Python anyway and wants workflow scaffolding around it — not a canvas to drag.

Self-hosted developer platform

Windmill. Internal tools + workflows + scripts in one self-hostable platform. Heavier than n8n, more code-native.

No-code visual canvas

Make. Cleanest canvas for non-technical operators on light-to-medium volume. Per-operation pricing bites at scale.

Quick prototype / niche SaaS glue

Zapier. Fastest first win and the broadest integration catalog. Outgrown the moment automations become load-bearing.

Open-source n8n alternative

Activepieces. True OSI-licensed self-host with a Make-like canvas. Smaller ecosystem, friendlier license.

Human-in-the-loop SaaS

Relay.app. Workflows that explicitly pause for human review. Niche, but the right shape for some compliance-bound flows.

Migration paths

A workflow system you can't move off of isn't infrastructure — it's a SaaS hostage situation. Every platform here gets compared on how it accepts inbound and outbound migrations.

Operational realities (every platform answers these)

If a workflow is going to run unattended in production, these are the questions that quietly decide the platform for you.

Workflow ownership
Can you export, version, and rebuild your workflows somewhere else — or does the vendor own the artifact?
Self-hosting + data residency
Does customer data ever traverse a third-party SaaS workflow runtime?
Cost model under complexity
Does adding a step multiply the bill, or is the unit “execution”, not “task”?
Error paths and observability
What happens when a workflow fails at 03:00, and who finds out?
Code escape hatch
When the canvas runs out, can you drop into JS or Python without leaving the workflow?
Access control + audit
Can ops, engineering, and external reviewers each see what they need — and nothing else?