AI Agent that builds deterministic workflows

I work in a small company and have always enjoyed automating workflows for myself and my team, starting out with macros many years ago, then python and then in the last year, agents. But while it's been fun to play around with, most of the stuff that needs to be automated doesn't actually need an agent on each run. It's deterministic stuff, too tedious or time consuming to do manually each time, maybe with 1 or 2 judgements needed along the way but the process is always the same. So over the last months we've been experimenting with our own "automation platform", which is a combination of an AI agent that builds the automations (using Claude or ChatGPT subscriptions so no token cost to worry about), and a "workflow and runtime management layer", that catches errors, tracks issues, manages credentials etc. The agent writes the automations in a very specific structure where it tests everything along the way, writes input and output validation per step, documents how it works etc. If a step fails the platform will catch it and alert the owner for review, and if the agent needs to fix something it will analyse the issue, propose a plan and get approval before doing anything, instead of going off on its own. A typical automation build uses 2-3x the amount of tokens as just having an AI agent do it once, but after that it's either 100% code, or maybe with a few simple AI steps that cost a couple thousand tokens on a cheap model. We think it would be cool to share with others, get some feedback on it and build something we can use for more and more things, but because it's not just a harness it needs a bit of polish before we push it out there. So to judge how to prioritise it I wanted to ask if anyone else have done something similar? Is it something you think could be useful - and what would be your concerns?