.webp)
I doubt I'm the first to propose this concept, but I'm probably the first to name it.
Drifting towards determinism (DriDe - as in "DRY'd" - Don't repeat yourself) is what everyone will do in 2 years, and I'm telling you to start today.
And if you want a one-size-fits-all explanation: I propose that while most people try to add more AI and then protect it, we should build systems in a way that removes AI from them completely or as much as possible over time.
Ok, smart guy: what is the drift towards determinism?
Well, it's not the long-awaited second installment of Fast and The Furious 3 (unfortunately, but it would be an amazing title, right?).
No, it's a philosophy about how we should think about AI.
Everyone is using AI agent systems and burning tokens as if they will run out one day.
Watching people spend $20 to set a reminder to buy milk hurts my soul (yes, really, that happened... heartbeat every 30 minutes to check the time was eating tokens!).
I think we will soon wake up from this "everything is solved by AI" fever dream and realize that there is a simple flow that will allow us to do almost anything, at a fraction of the cost and environmental impact.
Simple steps:
- Give the AI agent system a "novel" task that it hasn't seen before and let it burn a large amount of tokens to solve it.
- Put a second agent system at the end that looks at what could have been solved deterministically (i.e. in code).
- Build the tools for repeatable parts.
- Next time a similar task is presented, introduce the tools in step 1.
- Let's see if we always use tool1 and feed it to tool6 - connect them.
- Repeat the process until you write as many parts of the AI as possible.
- There are many small nuances, such as turning to AI if a tool does not provide the correct result for this work, running shadow versions of workflows to verify that we are really improving, providing feedback on the final result to fine-tune it, producing a system that the LLM can understand, full process tracking... but that can be solved correctly :-)
Over time, your non-deterministic AI-powered workflows, which cost $50 to run and only work 50% of the time without you asking them again, become glorious automations that use $0.02 worth of AI to categorize them and then simply run them in code.
It's faster, more consistent and can be trusted.
This is where we are headed.
Yes, people are developing skills and tools, what's new here?
That's the point: we already have the things we need to make this work, but we fundamentally get it wrong in how we treat them.
No one, and I mean no one, has a goal of removing AI from a process they are currently doing with AI.
Name me a tool/product that has used AI the least in the last year.
Go ahead, I'm waiting.
And yet, that is really what I propose.
AI is used to outline the outline of a process. It's expensive and slow (compared to code), but it solves a repetitive business process problem.
Then, you analyze that process. Do I really need to pass all 12000 rows of our company's customer list to AI to know who to call next? No, a simple tool to capture the next 5 people who haven't been called in a month.
Do I even need to give that tool to the agent? No, it should be made part of the context so that you have that information and we save a lot of round trips.
Wait a minute, are we giving AI a tool to then search for your website? Well, if you need that information, we should do it automatically and incorporate it into the context.
Wait a minute, have we scanned your website before? Do we have the information? Do we even need the agent to activate?
You get the idea.
Crystallization is the key
Every time you call an AI, you literally roll the dice.
It has improved a lot, but it is and always will be a no-
