Andrey Golub

The classification instinct.

By education I am a mathematician- tensor calculus and graph theory back when neither had a market value. By career I am an AI architect and R&D leader with twenty-five years, twenty-plus patents, and a standing practice of being brought into initiatives where probabilistic AI must meet symbolic backbones, contracts, and auditability. The industry is now rediscovering my research territory as if it were new. I write these notes to show that the instinct to classify first and build second still produces better systems than the instinct to generate first and explain never.

The Era of the Afternoon Agent Is Closing.

The level nobody is selling…

Decisions about enterprise AI are made at four levels.

1. The model level. Which model, which prompt, which fine-tune, which context window, which benchmark.
2. The tooling level. Which framework, which vector store, which orchestrator, which evaluation harness.
3. The solution architecture level. Which patterns, which boundaries, which contracts, where decision authority sits, what the system is allowed to do without asking.
4. The portfolio level. Which use cases, in which order, against which return.

The industry has extraordinary depth at levels one, two and four: benchmarks, leaderboards, framework comparisons, maturity models, ROI calculators, a whole consulting economy built on sequencing use cases.

Level three is thin to the point of absence. Not because it is difficult. Because it does not sell. Nobody ships a boundary. Nobody demos a contract. There is no launch event for a well-placed validation gate, and no analyst quadrant for knowing where authority sits. A boundary is invisible when it works, which is exactly what makes it valuable and exactly what makes it unmarketable.

And that is where the failures land.

Read the post-mortems of the last two years and the stated causes are almost always level one or level two. The model hallucinated. The framework was immature. The retrieval was poor. Those are real observations, and almost never the actual failure. The actual failure is that nobody named where the authority sat the decision, so when the output was wrong there was no boundary at which to catch it and no artifact to inspect afterwards. A system without named internal structure cannot fail in a diagnosable way. It can only disappoint.

Solution architecture is the discipline of putting that structure in before it is needed- deciding what the model may determine and what it may only propose. In classical enterprise IT this level is well staffed. In AI it has been treated as an implementation detail to sort out after the demo lands.

What changed?

I assumed for a long time that this would correct itself: that the wrapper phase would end on its own, that bounded systems with the model as a component rather than the brain would win because they simply work better, and that the market would arrive at the right architectures by ordinary evolutionary pressure.

It did not happen. Volume won. An assistant assembled in an afternoon kept being sold as a business function and bought as one. So the correction is arriving from regulation instead- not a sentence I write with any enthusiasm.

An AI system that interacts with people in Europe must now be able to say what it is and on whose behalf, it is acting. That reads like a disclosure requirement. It is not. Read structurally, it decomposes into five capabilities that must exist before any label can be attached: the system must be a nameable thing with declared purpose and limits; it must disclose the principal it acts for, which presupposes a real delegation chain, not authority implied by a prompt; it must have named moments where notice is due, which presupposes gates, not a trajectory; it must carry that obligation across handoffs, making disclosure a protocol property, not an interface string; and it must mark its output as synthetic at production time, making provenance a construction property, not a label applied afterwards.

None of these five are compliance features. They are the ordinary properties of a governable system: declared scope, explicit authority, materialized decision boundaries, traceable handoffs, provenance by construction. The regulation is not asking for a banner. It is asking whether the system has internal structure that can be pointed at. You cannot label what you cannot locate.

This is why the era of the afternoon agent is closing- not because the industry grew wiser or the demos stopped being impressive, but because the difference between a system with named internal structure and one without has stopped being a question of quality and become a question of admissibility.

I came through the last three years without becoming a generator of LLM automations, and it was closer than I would like. Several capable architects I respect did not: they spent two years shipping wrappers and are now rediscovering, with some irritation, that they knew how to do this properly all along. They are about to be useful again, which is the most encouraging development in this field in some years.

What a pattern actually is?

If level three is the missing level, the natural question is what it is made of. The answer is patterns- a word that has been used so loosely it barely means anything. Four different objects are routinely called patterns, and none of them is one.

The template is a structure with blanks in it. It tells you what to fill in and never why this shape rather than another; it has no alternative, which is the tell. The best practice is a recommendation with survivorship bias attached: it tells you what worked somewhere, not which conditions made it work, so it cannot tell you whether your conditions match. The reference architecture is a complete stack presented as neutral, every box a decision already made with the decision itself deleted from the drawing. The vendor blueprint is a reference architecture with particular products in the boxes, rendered in the visual language of a law of nature.

What these four share is the interesting part. None of them contains a question. They contain answers with the question removed.

A pattern is three things at once. It is a named recurring structure: something observed more than once, in more than one context, carrying a name that survives outside the room where it was coined. It carries a discriminating question: the one whose answer tells you whether you are inside this pattern or the one next to it. And it states a trade-off: what you give up by choosing it. A pattern with no cost is not a pattern, it is an advertisement.

The discriminating question is the load-bearing element. Remove it and the whole thing collapses into one of the four impostors above. Keep it and the pattern stops being a thing you adopt and becomes a thing you can test yourself against.

Take hybrid search. The market definition is keyword plus vector, fused- a template, and it is why so many teams implement it and then cannot explain why their retrieval is still wrong. The discriminating question is different: what kind of evidence is being combined, and at which stage? Ask it and the single label separates into distinct retrieval patterns- lexical and vector fusion is only the first- each combining a different class of evidence for a different reason, with different failure modes. A team that knows which one it is building can debug it. A team that has implemented "hybrid search" can only tune it.

Take the AI agent. The market treats it as one category with a maturity spectrum inside. The discriminating question is where decision authority sits. Ask it and the category separates into distinct classes, and the three critiques of agentic AI now circulating- that anthropomorphic framing degrades human oversight, that the return on autonomous agents does not materialize, that outsourcing execution to a frontier model transfers control to its owner- turn out to be three symptoms of one root. All three follow from the model, owning the decision. None follows from using a model. Which gives the shortest, useful diagnostic I know remove the model, and ask whether the process still exists. If yes, the model is a component and the system is governable. If no, the model is the system, and every critique above applies to you.

Patterns, not tokens!

Which brings me to the name of this series, and the distinction underneath it. Tokens are generated. Patterns are identified.

Generation produces something plausible from a distribution over what has been written before. It is an extraordinary capability and I use it daily. Identification produces a commitment: this is the case, that is not, and these consequences hold. Different operations- and the failure mode of the past three years has been using generation where identification was required. A model can generate a fluent account of why a system failed. It cannot tell you which question was never asked, because that question was never in the training distribution- it was never asked by anyone.

This is not an argument against models. It is an argument about placement: hybrid AI, where probabilistic components sit inside structures that are not probabilistic, bounded by contracts, validated against symbolic backbones, auditable after the fact. Ontologies, knowledge graphs, context graphs, semantic layers- the unglamorous machinery that makes the glamorous part safe to deploy.

Classifying first and building second is not pedantry. It is the only reliable way to know what you are building. A mathematician learns early that most problems dissolve once you have the right decomposition, and that no amount of computation rescues the wrong one.

Why do I write these?

I do not write these for reach; the number attached to a piece of writing has no relationship to whether the writing was right. I write because classification is how I think, and because writing forces classification to be honest in a way thinking alone does not.

What follows is one overloaded term at a time, taken apart and put back together with the precision it deserves. Knowledge layers, semantic architecture, retrieval and evidence, reasoning as a distributed capability, where authority sits in agentic systems, what can be proven after the fact. Each has the same shape: a term the market treats as one thing, a question that separates from it, and the consequences of knowing which one you are in.

An afternoon is enough to assemble an agent.
It is not enough to name a boundary.
Regulation has just made that difference load bearing.

#AI #EnterpriseAI #AIArchitecture #AIGovernance #AIAct #HybridAI #AIAgents #SolutionArchitecture