Graph engineering is workflow with a new sticker

It feels like every tech article I come across lately is buzzing about the new Graph Engineering concept.

If you’ve been working in tech for a while, you’re probably having massive flashbacks right now.

We didn't call it graph engineering. We called it workflow.

For years I worked on Enterprise Business Process Management (BPM) like IBM FileNet Content Engine and Process Engine. And here's what a process map looked like:

Steps that do work. Route conditions that decide what happens next. Data fields flowing through the process. State persisted in a database so a workflow could survive a server restart. Queues where a human had to review and approve before anything important moved forward. Deadlines, timers, escalation paths, exception maps.

Now read how graph engineering is described in 2026:

Nodes that do work, conditional edges for routing, a shared state object, checkpointing for durability, human-in-the-loop interrupts, error recovery paths.

Same machine. New sticker.

So what's actually new?

In the old days, every single route and rule was written ahead of time. The workflow could never surprise you. In an agent graph, there's a reasoning model inside the node deciding where to go next, at runtime.

It's a state machine where some of the states can think.

And that's exactly why the old machinery is being rediscovered. Checkpoints, approval gates, timeouts - they weren't bureaucracy. They were safety nets. We built them for unreliable humans. Turns out they work just as well for unreliable models.

The BPM engineers weren't behind the curve. They were just early to the party.