Workflow Automation: Letting AI Write Workflow Code

Workflow automation is one those perennial technology efficiency promises that has never quite delivered but never goes away.

The appeal of the idea is obvious, being able to easily direct computers to perform process tasks without deep technical expertise is the holy grail of the promise of computers. The excitement around AI agents is the latest incarnation of this promise.

Workflow automation can be implemented in many forms, the most common is some sort of visual composition approach. Drag-n-drop workflow builders for instance are quite common.

The fundamental problem that workflow automation has faced is that it attempts to allow non-programmers to...well, program. This is a fundamental contradiction that always reveals itself once one tries to use these products to do anything beyond the demo reel.

More recent takes on the idea have taken on the problem with a hybrid approach, ie code when you need it and UI when you don’t,...this is the literal tagline that the popular workflow automation product n8n uses. We believe this hues closer to the right approach for delivering on the promise of workflow automation; of course it means that the old promise of not needing to know how to code will have to be tabled.

In addition to the hybrid approach, there is now technology that may finally deliver on the promise of workflow automation and that is AI CodeGen. The myriad capabilities of GenAI can be leveraged to address a lot of the short comings of conventional workflow automation.

GenAI’s primary capability is the ability to take fuzzy input as free-form information (text,audio,image) and make sense of it. This is essentially what workflow automation products have attempted to do with all those visual mnemonics intended to be substitutes for writing code.

For existing products, GenAI can help with filling in the gaps between visual artifacts and what a user actually needs, solutions that combine code and non-code elements will do well with GenAI essentially filling in the blank for the code part of workflows.

For greenfield products, we think doing away with the usual drag-n-drop or configuration based approach is best. In other words just let GenAI write the workflow code using the tools you make available to the code writing agent.

This is the approach we’ve taken with our workflow automation capability, nothing to drag-n-drop. The downside is that if you need to do manual changes, you have to alter the AI generated code.

We are leveraging CodeGen tool use, which is just telling the AI about the required tool API and having it use that to wrap logic based on user requirements. In this context, ***tools ***are the usual integrations that form the building blocks of conventional workflow solutions.

A demo of AI writing workflow code: