I think it's the cost (both tokens and initial setup and debugging the workflow).

It's really easy to spend a gazillion tokens by mistake and the resulting workflows are still probabilistic and open for abuse or failure.

Maybe it's too soon, we need more time to figure out what works and what doesn't.

We're still looking for the right balance for when a probabilistic solution (i.e AI) works and when we need a deterministic one (classic software).

Correct and I think for non technical folks just creating pure ai agents is much easier than creating these workflows which can be harder to test as well.

This will come with higher risks, no transparency, and higher costs, of course. That seems like a recipe for even higher frustration than normal software incurs.

My guess is that except for simple one-offs or low risk situations, any recurrent use of AI will require some sort of expert to handle it, at least for a while.

Liability plays a big role here, in the end a person needs to be made responsible for the output of stuff and face consequences.

There does seem to be a need for a semi-deterministic workflow where nodes in the workflow are agentic, i.e., powered by LLMs, but the workflow itself remains deterministic. Lets say my daily workflow is to download a PDF from some website, extract some information from it, and email that information. I cannot just ask an agent to script this workflow, because the website will keep changing, the PDF layout will keep changing, etc. I need a "download agent", an "extract info" agent, and so on which will examine the website, PDF, etc. and generate any one-off scripts everytime the workflow is run.

My solution for that is to use skills and add a omnibus shell script in each particular skill folder that does all the deterministic steps and then write the skill markdown to only use the script, usually passing a command line step parameter to the script. I also usually add a note in the skill markdown that it should monitor how it calls the script and suggest any improvements when the skill is used.

As an example I have a `today` skill that has a today.sh script that can check my email, bug tracker, calendars, custom timesheet system and other systems via the script and then the skill assembles a nice summary of what is happening today along with a suggested schedule.

This is exactly what I’ve been building with ctx.traits :)

https://ctx.company/traits/docs/

I guess the tradeoff is doing it in a workflow builder vs just dumping all your requirements in a single agent, With LLMs getting smarter, they can create sub agents automatically if required and take the abstraction away from the user.

And without an LLM you wouldn't even try?

I don't mind semi-deterministic for plenty of things. But these agents like hermes, can codify a skill to be deterministic.

n8n will do that

I‘m wondering what that means for products like retool. I used it a lot in the past for internal tooling but it’s now a a segment where coding agents are likely to be a better fit, and the AI integration within retool is messy compared to just letting Claude create the web app

The problem that workflow engines solve (configurable code) is just better solve with agents just maintaining the code.

2 years ago (during the agentic workflow boom) prompting was annoying and tedious. Context was small, harnesses were weak. You needed to string along multiple complex prompts to get anything done.

Now, you can give an agent 50+ tasks and it will churn through them.

Because it is a workflow builder. It is the wrong mental model. Agents and workflows are completely different things.