Contextual irony aside, this is a big reason why the proposal of leveraging AI agents for workflow processing in lieu of using them to develop fixed software to perform the same functions has always struck me as weird, and of late come across as completely nonsensical.
If you're paying someone else to run the inference for these models, or even to build these models, then you're ultimately relying on their specific preferences for which tools, brands, products, companies, and integrations they prefer, not necessarily what you need or want. If and when they deprecate the model your agentic workflow is built on, you now have to rebuild and re-validate it on whatever the new model is. Even if you go out of your way to run things entirely locally with expensive inference kit and a full security harness to keep things in check, you could spend a lot less just having it vomit up some slopcode that one of your human specialists can validate and massage into perpetual functionality before walling it off on a VM or container somewhere for the next twenty years.
The more you're outsourcing workflows wholesale to these bots, the more you're making yourself vulnerable to the business objectives of whoever hosts and builds those bots. If you're just using it as a slop machine to get you the software you want and that IT can support indefinitely, then you're going to be much better off in the long run.
It's the siren song of the myopically lazy. It's laziness today in exchange for harder work tomorrow, with the wager that tomorrow's harder work will be even lazier thanks to advances in technology.
Whereas I'd self-describe as "strategically lazy". It's building iterable code and repeatable processes today, so I can be lazy far into the future. It's engineering solutions today that are easier to support with lazier efforts tomorrow, regardless of whether things improve or get worse.
Building processes around agents predicated on a specific model is myopically lazy, because you'll be rebuilding and debugging that entire setup next year when your chosen agent is deprecated or retired. Those of us building documented code with agents today, will have an easier time debugging it in the future because the hard work is already done.
Incidentally, we'll also have gainful employment tomorrow by un-fucking agent-based workflows that didn't translate into software when tokens were cheap and subsidized by VCs for market capture purposes.
If you're paying someone else to run the inference for these models, or even to build these models, then you're ultimately relying on their specific preferences for which tools, brands, products, companies, and integrations they prefer, not necessarily what you need or want. If and when they deprecate the model your agentic workflow is built on, you now have to rebuild and re-validate it on whatever the new model is. Even if you go out of your way to run things entirely locally with expensive inference kit and a full security harness to keep things in check, you could spend a lot less just having it vomit up some slopcode that one of your human specialists can validate and massage into perpetual functionality before walling it off on a VM or container somewhere for the next twenty years.
The more you're outsourcing workflows wholesale to these bots, the more you're making yourself vulnerable to the business objectives of whoever hosts and builds those bots. If you're just using it as a slop machine to get you the software you want and that IT can support indefinitely, then you're going to be much better off in the long run.