Cloudflare's docs say it can't be done: Workflows don't run on Workers for Platforms. So I rebuilt the Workflows API myself.
Grasp rebuilds a company's processes so AI does part of the work. I build the platform those processes run on, on Cloudflare.
1/6
So I recreated the Workflows API to run inside any worker, normal or Workers for Platforms, with the same shape. It's called Workerflow.
Now isolation doesn't limit what a client's processes can do. They keep running, and keep saving the team time.
5/6
It's the same problem I worked on for six years: a process has to run reliably, or the time it saves goes into fixing it. Only the author changed.
Engineers: do you let agents write production code yet? Where do you draw the line?
5/5
I was hired as a no-code developer. I spent the next six years writing code.
At Betty Blocks, a low-code platform, I built the workflow steps clients needed, the parts the standard blocks couldn't do.
1/5
The types don't allow an agent to approve a payment, so the check fails before anything runs, and the agent hands the approval to the controller.
After that, every workflow is tested on past work, and the person who owns the process approves it.
4/5
Any run can be traced back to a person or an agent.
That's what keeps the time you save actually saved. Nobody has to spend their week double-checking the AI.
How do you keep track of what AI does in your systems today? Curious how other teams handle it.
5/5
If an AI agent changes something in your CRM at 3 a.m., can you tell who allowed it?
You should be able to, especially once AI isn't just writing drafts but doing part of the actual work.
1/5
If AI calls can happen anywhere in your systems, finding that out means searching through all of them. So in Grasp every call to an AI model goes through one gateway, where it's checked and logged, and every action is logged as well.
4/5
Today Grasp comes out of stealth. We make companies of 50 to 500 people AI-native, by implementing AI inside their actual business processes.
I've spent the last months building the platform underneath it.
1/5
4. An agent can only do what an admin allowed it to do.
5. Every call to an AI model goes through one gateway, and gets logged.
None of this shows up in a demo. It's what makes it possible to trust AI with real work.
4/5