Andrey Gaag

The Canon Has Not Yet Formed

The industry is very young

In many fields of human activity, practices and standards have been refined for centuries or even millennia. Software development has existed for less than a century. Only eight decades have passed since it emerged in the 1940s, and the industry has already undergone several shifts that changed both its practices and its standards:

Fundamental shifts are accelerating

Our industry has always been turbulent. But fundamental changes that once took decades now fit into a few years, and the pace does not appear to be slowing.

The principles do not change

The fundamental principles and problems remain with us. Managing complexity - the challenge Uncle Bob taught us to confront - was, is, and will remain the central task from assembly language to LLM agents. Abstractions save us, yet they still leak: hidden details slip into other layers or protrude through their boundaries regardless of the stack, target device, or AI-first label in a product description.

Data persistence and consistency have not disappeared either. Our beloved ACID still matters: data is everything, and without it a program loses its purpose. Networks remain unreliable whether an AWS outage cuts off an LLM or a website becomes unreachable in the 1990s. Coordination under concurrency has not gone away: just as two processes need synchronization when sharing memory, two subagents need coordination when editing the same file. We could name a dozen more examples, but the point would remain the same.

Tools change speed, not the physics of complexity.

Each new era has moved and encapsulated complexity rather than eliminating it, building new abstractions along the way. Despite the marketing and panic around the technology, LLMs do not invalidate computer science. They turn part of deterministic engineering into probabilistic engineering.

Peak turbulence

In production-grade LLM agents, all this turbulence reaches maximum intensity: tooling and conceptual shifts happen within a few months. Fundamental principles and established methods help to a degree, and the industry is gradually crystallizing new practices, but in the most interesting areas we still have to feel our way through the dark.

Anthropic may describe some of its hard-won lessons on its blog, but it will not give away durable practices that create a competitive advantage. Google engineers published Agentic Design Patterns, but without testing complex combinations of those patterns in practice, the book is less useful than it could be. Humanity already has some working knowledge, but it is scattered across the minds of individual engineers and teams. Chip Huyen confirms this in her book: she had to gather information piece by piece through private conversations with engineers.

We are helping form the canon

The engineering discipline of agentic systems is taking shape right now. Today’s experiments and mistakes will become tomorrow’s standards, and we are participants in that process. There is no canon yet, nor are there stable frameworks proven over years and millions of user scenarios.

We must keep watching the industry, learn from the best, absorb the rare pieces of systematic work we can find, and, above all, practice: try new models and architectures, build agents, and test ideas against reality.

We live in a remarkable time - a time of discovery, change, and opportunity. This new engineering discipline is forming before our eyes, shaped not only by technology giants but also by small teams and even individuals. It is all in our hands and minds.