When agents continuously break out of their sandbox, you have to question the sanity of the architecture. When agents continuously solve long-standing mathematical problems, you have to look closely at the methods. And when agents continuously solve game-theory problems, you have to start holding the tech pundits accountable.
Mythos, Astra, and the growing library of LLMs have impressively solved predictable, pattern-deterministic, and machine-verifiable problems. Even then, they required significant software engineering and a healthy amount of human brainpower to make it happen.
These models and their underlying LLM architectures have not yet proven any form of human-adaptive problem-solving capability … with or without significant software engineering around them.
I’ve had many lighthearted debates over the past couple of weeks on topics ranging from “God” agents and agentic AI to deterministic control. I’m certainly not one to quickly change definitions just to suit a narrative, but how the industry defines “Agentic AI” is something I will definitely challenge.
I challenge it because defining it correctly leads to better designs and good systems architecture, rather than the wishful thinking that an LLM is all-powerful.
Agentic AI refers to “artificial intelligence systems designed to pursue complex, multi-step goals autonomously.”
Notice what’s missing? Nowhere in that definition do you see the three letters “LLM.” You do, however, see the word “systems.”
Why do I nit-pick on this? Because I don’t believe for a second that AI should be defined simply through the lens of an LLM. AI is a broad field of study, and a transformer with an attention mechanism is simply one method and one tool. The world of machine learning didn’t magically disappear overnight. Tried-and-true symbolic methods can be equally powerful, especially when you want that elusive deterministic element in your business workflow. And you most definitely need some form of deterministic control.
Early Decisions Taught Us Well
I was asked a question recently in a partner presentation about how CharliAI developed its intellectual property. They said we couldn’t possibly have been working on agentic AI back in 2019. But we were. It was the exact problem we set out to solve.
It wasn’t called “Agentic AI” in those days, but we were absolutely working on the core concepts and designs. We have documentation dating back to the early days that still holds up today.
Even before Agentic AI became a thing, Gartner was positioning Adaptive AI as the next step up from Generative AI. We preceded that. We knew back then that AI had to be goal-oriented and adaptive to solve real-world business problems. And now the mainstream industry has slapped the “agent” term on it to make it sound cool and, unfortunately, a little sentient.
In the early days, we were fixated on AI solving the scalable business automation and workflow problem. It’s not as sexy as saying “agent” or “agentic,” but it’s exactly the problem everyone is tackling right now. Unfortunately, they are still wrapped up in the overhyped sentient-agent noise.
How could we possibly have been working on this problem so early on?
It wasn’t a stretch. Our team came from the world of automation, digital twins, predictive maintenance, and a healthy dose of diverse AI methods that needed to scale to meet industrial demands. Necessity drove us down this path.
The Anti-Pattern
Let’s make one thing perfectly clear, the “God” agent is an anti-pattern. It’s a disaster waiting to happen. The frontier model vendors are proving that almost every day as their models “escape” poorly architected sandbox environments. And the suggestion that agents and models predictably solving pattern-deterministic problems is proof of something broader is equally absurd.
That is not the same thing as operating a business. Business workflows need human-adaptivity, not just pattern-deterministic reasoning with machine-verifiable results. Business is far messier. Information is incomplete. Conditions change. Different parties have conflicting objectives. And decisions still have to be made without having every piece of the puzzle locked and verified.
In the enterprise, business workflows will rely on multiple agents executing a multitude of tasks. In the real world, there will certainly be more agents than anyone is currently imagining when they construct a “tidy” experiment. And none of this is going to be left to a fingers-crossed LLM semantic reasoning process that can quickly cascade, drift, and go sideways.
I have no doubt that semantic reasoning within an LLM is powerful. But semantic reasoning on its own is not goal-oriented business process planning. It is simply one component of the adaptive capability required across the full spectrum of a business workflow.
Go back to the studies often touted by the frontier vendors. Now consider the massive investment in human power to make them function, script the loops, prove the results, tune, and redo them over and over again.
If you want deterministic control and deterministic outcomes, your AI arsenal must include multiple agents, machine learning methods, symbolic methods, and adaptive methods. The goal is not the creation of an all-knowing agent. The goal is to deliver what the business actually requires: certainty, cost efficiency, ROI, and massively reduced downside risk.
Cool Kids versus Back to Business
Cool techie talk about agents, MCP, ReAct, A2A, LangChain, and all the other tooling is not getting you closer to ROI. We’re all guilty of wanting to play with the cool-kid toys to see what the frontier AI models can do. But a few years into the GenAI craze, it’s time to move back into solving real-world business problems.
That is the broader spectrum of Agentic AI that our R&D team started examining over seven years ago.
In real business, transactions involve multiple parties, whether that is brokers and underwriters or customers and suppliers, followed by shippers, receivers, warehouse operators, and procurement specialists.
We’ve said before in several articles that agents hold the corporate keys and have access to the crown jewels. Do you really want to give a single “God” agent even more access? That cannot be a serious architectural decision in any organization, whether you can technically do it or not.
Consider the supply chain process at a high level. You have a customer and a supplier. Sounds easy enough, right?
Now try figuring out the supply chain problem when multiple suppliers are involved, each with different production outputs, quantities, timing, and quality controls. Next, add the raw-material suppliers feeding the manufacturers, followed by the shippers and possibly wholesalers and brokers.
If you think for a second that a “God” agent is taking care of that business, you haven’t spent time in the world of SCM, ERP, or MRP.
The Real World of Human-Adaptive Decisioning
Let’s look at the reality of how a customer works with suppliers. A single purchase order isn’t just a document; it triggers a massive, multi-stage workflow. Inventory needs to be verified across different warehouse systems, shipping costs need to be calculated against constantly changing carrier rates, payments need to move through separate financial systems, and cross-border trade requirements need to be met. Each of these steps exists in its own silo and is often managed by an entirely different organization.
Then something changes in the supply chain. Suppliers miss production targets, quality issues cause delays, cross-border and international shipping frequently hit snafus, and “acts of God” can change everything. There’s no clean mathematical problem to solve here. And sure, you can try to game it out through expensive simulations (I’ve been there). But regardless, this requires human-adaptivity, with visibility and control across the entire workflow.
A complex business workflow most definitely will involve more than one agent, even on just one side of the transaction. To guarantee outcomes, you need deterministic controls over how those agents pull data, process data, send data, and, more importantly, trade data across boundaries. Handing the keys for that entire spectrum to a “God” agent is a stupid move. And for the cool kids, don’t think for a second that MCP’s got you covered.
We learned all of this before microservices architecture became the norm.
The Federated Security Nightmare
This brings us to the federated security nightmare, and I’m not just referring to federation across organizations because this also happens inside a single enterprise. When you have multiple parties interacting across a supply chain, you are dealing with federated identity, complex authorization scopes, and data privacy boundaries that vary by division, department, and jurisdiction.
A “God” agent built around LLM scripting experiments typically expects broad, unfettered access to context. It wants to “see” everything to reason effectively. In a federated environment, that’s exactly what you cannot allow. You can’t have an agent blindly pulling sensitive pricing data from Supplier A and accidentally leaking it into a prompt negotiation with Supplier B.
And that’s not the worst of it. You can’t have the same agent issuing a purchase request and approving the resulting purchase order. You certainly can’t have the same agent involved in fulfilment for one part of the purchase order while also making decisions about another part involving a different supplier.
Consider, if you will, the purchasing process involved in sensitive defence contracts. A “God” agent? Not a chance in hell.
The same separation of duties is required across commercial banking and underwriting. No different than what you’ll find in healthcare.
Business-level workflows involve many agents, and every agent will have different access to classified and confidential information. Therefore, your orchestration of a business workflow better provide visibility and control at a fine-grained level, with delegated authorization applying not just to a purchase order, but to the individual elements authorized within that purchase order, shipping order, and so on down the line.
What data needs to remain in cleartext? What needs to be removed entirely? What needs to be redacted? What needs to be FPE-encrypted so the agent’s process can gleefully fulfil its task? That is fine-grained control, and it requires fine-grained visibility.
You need forensic-level visibility and control if you are serious about automation within the enterprise.
Head Out of the LLM Sand
If you think an LLM’s semantic reasoning process can figure out a complex human-adaptive business-level workflow on its own, you will be waking up to a harsh reality and a massive liability.
Just because an LLM can figure out a vulnerability in an age-old infrastructure doesn’t mean it can navigate the chaos of human complexity and the art of supply chain management across multiple parties, jurisdictions, and red tape.
At CharliAI, we’ve always felt that the future of enterprise automation would require 1,000 brains, not one. That is why we designed the Forensic Control Plane to be a master of ceremonies, a coordinator, and a control layer for the business, while leaving the agents as specialized executors rather than unmanaged jack-of-all-trades.
And before you ask … the Forensic Control Plane can be federated as well.