for simple workflows, smart agents can effectively run them as deterministic workflowsfor dim - workflow complexity -- simple, smart agents can effectively run them as deterministic workflows (effectively deterministic)
I've been thinking about something wrong.
I've been thinking: AI is not deterministic, so it doesn't make sense to use it in deterministic workflows
But that's not true, actually. It really depends how hard the workflow is.
For example, I'm working on a pretty simple workflow right now. I'm developing an agent skill that does application releases. There are a bunch of bash scripts in the project that are used for stuff like compiling the application, syncing it to S3, etc. To properly do a release, I need to essentially answer all these questions.
- should it use the slow (but needed-for-real-customers) Apple notarization process when building the app?
- which destination bucket (dev or prod)
- do I want to include release notes? (I might skip it if I'm just doing a test)
- should it increment the
auto-update-version.txtthat causes the app to ask the user if it can auto-update?
In the past I absolutely would have created a release shell script that wrapped all the sub-scripts and forced you to answer all those questions at the start.
But then I stepped back and thought about it. This is simple enough that Claude Code running AI unit - claude 4.6 Opus is never going to screw it up, and by making it a text-based skill, I can do all sorts of flexible stuff like have it help me draft the release notes, or ask it to name the release something really weird for a very specific purpose.
The other thing that's nice is that I can run it with no options and (like shown in example - Claude Code Agent Skill - AskUserQuestion), it will ask me all those questions...
... but also, I can use wispr flow to just describe what I want beforehand so those questions will be pre-answered. If I happened to forget to specify a single choice ahead-of-time, it will bug me to answer just that one question. Fantastic!
So yeah, I think for a workflow like this, Claude Code is effectively deterministic. Hmmm... that's an update to my priors