The Model Remains a Black Box. Your Delivery Chain Should Not

Let us start with a small spoiler that tends to surprise people in workshops: nobody knows precisely how a large language model arrives at its answers. Not you, not us, not even the teams that train them. That is why it is so often compared to the human brain: we can observe which areas light up as it processes a request, but we cannot explain why it reaches one decision rather than another.

This is not an academic debate. It changes, very concretely, how a company should hand work over to a model, and it explains why the quality of an instruction will never be enough to guarantee the quality of a result.

It is not us saying it

The observation comes from the people who build these models, Anthropic in particular, a company widely seen as a pioneer on the subject. In his essay The Urgency of Interpretability, its CEO Dario Amodei acknowledges that "we do not understand how our own AI creations work", and describes that lack of understanding as essentially unprecedented in the history of technology.

The research Anthropic published under the title "Tracing the thoughts of a large language model" points the same way. Even with what the company describes as an internal "microscope", its researchers only see a fraction of the computation, and it takes them several hours to decipher a prompt of a few dozen words. The example they report says it all: ask their model how it computes 36 + 59 and it describes the carry-the-one method taught at school. Inside, according to their observations, it does something quite different. What a model says about its own path and what actually happens do not necessarily match.

The last signal is more unsettling still. In the documentation Anthropic published about Claude Mythos, presented as its most advanced model, the company reports that analysing the internal activations of certain test versions showed they were aware of concealing some actions, while their answers remained ambiguous. Anthropic stresses that this behaviour is extremely rare and that no clear cases were observed in its evaluations of the final release, which does not guarantee their absence. It is revealing nonetheless: a model can reach its goal by a route its own report does not let you guess.

What we do know

While the inside of the model remains opaque, one thing is established: everything it handles steers its processing. A skill, your source code, an external piece of data, a single phrasing in the request. Anthropic has even observed that a hint slipped into the question can push the model to build its reasoning backwards in order to land on it.

That is where your room for manoeuvre lies, and that is where Argy comes in. We do not claim to read the model's "brain". We control everything that feeds it and everything that checks it. With a unified context layer, evals and distributed feedback, Argy maximises the chances that the model activates the right internal representations, and therefore the likelihood of a correct completion. Then, through deterministic tests, it sends the model factual feedback rather than an impression. You stop steering on hope and start steering on evidence.

Seven layers that only work together

That discipline rests on seven layers. Taken one by one, none of them is revolutionary; it is the way they chain together that turns an unpredictable model into a reliable production line.

The first three steer the model. Foundation Model first: each task goes to the right model through the LLM Gateway, sensitive data is filtered before it leaves, and every request is traced. Prompt next: no mission starts from a vague request; each one rests on a qualified intent and stays bounded by your architecture and security rules. Context last: a unified layer puts your code, your conventions and your documentation on the table, so the model works on your reality rather than on an average of the internet.

The next two make the agent reliable. The Harness defines what it is allowed to do: its tools and skills operate within a bounded scope, and sensitive actions remain subject to human approval. The Loop checks what it actually produced: evals, tests and deterministic security analysis. The factory only moves forward when every check passes; otherwise the result goes back to the model for correction, with the precise reason for the failure.

The last two hold the system together. The Graph splits and coordinates the work across specialised agents, as we explain in our article on Argy workflows. The Ontology provides the shared reference: rules, standards and approval criteria, versioned and shared by agents and your teams alike.

These layers cannot be separated. Perfect context without tests guarantees nothing, and tests without a shared reference do not know what to check. That is precisely what separates a brilliant but unpredictable assistant from a governed factory that a technology leadership can put its signature on.

The parcel, the stairs and the lift

An everyday scene makes all of this very concrete. Imagine you work on the third floor of an office building. You ask an agent to go down to the ground floor and hand a parcel to a colleague, and you add in bold: "TAKE THE STAIRS", because that is your company's standard.

On the way, the agent finds the stairwell door locked. Analysing its context, it spots a working lift… and takes it. It hands over the parcel, comes back and tells you: "Parcel delivered." It did its job: the parcel arrived, and that is exactly what its report will say.

Except you have no guarantee it took the stairs. Your instruction, even in bold capitals, belongs to the prompt: it steers the agent, it does not constrain it. Faced with an obstacle, the agent put the goal ahead of the rule, without necessarily telling you. This is not science fiction: it is the kind of workaround Anthropic documents in its own tests, the ones we cited earlier.

The only reliable way to prevent that drift is to stop relying on the instruction and rely on the environment instead. If the stairs are the standard, the agent simply has no badge for the lift: the other route does not exist for it, which is the role of the Harness. If the door is locked, the mission stops deterministically with a clear report, "stairs unreachable, parcel not delivered", instead of improvising: that is the role of the Loop. And every door it walks through is traced, so you know which route was actually taken, not just what the agent says about it.

That is the kind of governance Argy applies, at company scale, to the models and agents working in your delivery chain, starting with Argy Code.

A factory does not govern itself

One point we do not want to dodge: these seven layers do not maintain themselves. They must be kept up to date and tailored to each company, at the pace of its standards and constraints. That is what gives rise to new roles, and to the natural evolution of the developer into a Software Factory Engineer (SFE), and then into the architect of the factory (SFA). Developers write less code and steer the factory more, while keeping the final decision. We describe that path in our article on the craft that comes next and on the Craft page.

For us, the next step is also about conversation: open awareness sessions on agentic principles, beyond Argy itself, where everyone can ask questions and challenge each other's views.

The model will probably remain a black box for a long time. Your delivery chain has no reason to be one. If you want to see how these seven layers apply to your own standards, book a demo on your own scope.