The fundamental promise of artificial intelligence has always been the reduction of friction between human intent and machine execution. We build these systems because we believe technology should eventually understand what we want to achieve without requiring us to painstakingly map out every how.
However, in our rush to make current-generation models “production-ready,” we have surrounded them with an elaborate web of external supports—memory layers, orchestration frameworks, and complex tool-calling wrappers. While these feel like permanent infrastructure today, they are often just placeholders for capabilities the models haven’t yet mastered.
If we look at the trajectory of the industry, the pattern is undeniable: whatever we build as a “harness” outside the model today becomes a native feature of the model tomorrow. Betting against the pace of model improvement is, historically speaking, a losing game.
The Divergence of Procedural and Preference Design
To build systems that don’t become obsolete with the next frontier model release, we must distinguish between two very different types of engineering.
The Fragility of Procedural Scaffolding
Much of what we currently call “AI Engineering” is actually procedural scaffolding. These are the rigid workflows where we tell the model exactly which steps to take: “First do A, then check B, then output C.”
This approach mirrors the “Bitter Lesson” described by computer scientist Rich Sutton. History shows that whenever researchers try to bake human-designed “cleverness” or specific heuristics into AI, those methods are eventually overtaken by general-purpose algorithms powered by massive scale and computation. Better models internalise these procedures. If your value proposition is simply a better multi-step prompt, that value will evaporate when the model learns to reason through those steps natively.
The Durability of Preference Harnesses
On the other side, we have preference harnesses. These don’t define the process; they define the standard. They represent the specific context of an organisation—its values, its definition of “good,” and its unique constraints.
A model might become smart enough to write a perfect legal brief, but it will never intuitively know your specific firm’s stylistic nuances or your particular client’s risk tolerance unless you tell it. These layers survive because they provide the signal the model needs to align its general intelligence with your specific intent.
Navigating the Practical Constraints of Today
While the long-term trend points toward total model absorption, engineering leaders in Bangalore and Mumbai still face immediate operational realities. We cannot ignore the “What” of our current technical environment:
- The Context Economy: Even as windows expand to millions of tokens, the cost and latency of processing massive histories remain significant. Until models can compress lifelong context natively and efficiently, external memory systems remain a necessary evil.
- The Determinism Gap: For high-stakes enterprise applications, raw models can still be unpredictable. Orchestration layers currently provide the guardrails required for compliance and reliability, though this gap narrows with every version update.
- Architectural Lock-in: Building entirely within a single provider’s ecosystem creates immense vendor risk. Smart teams are using thin, portable orchestration layers not because they love the complexity, but as a hedge against shifting pricing or API stability.
- Economic Realities: Frontier models are becoming cheaper at an incredible rate—roughly 75% year-on-year. This creates a strange paradox: it is often cheaper to wait for a model to get smarter than it is to hire an engineering team to build a complex workaround.
A Strategy for the Adaptive Builder
If the endgame is a collapse of the orchestration layer into the model itself, how should tech leads and founders build today? The goal is to ensure your system gets simpler as models get stronger.
Prioritise Model-Agnostic Intent
Invest your energy in documenting “what excellence looks like” for your specific use case. This means building clean, structured datasets of your preferences, edge cases, and success criteria. Whether you use a model from OpenAI, Anthropic, or an open-source alternative, this “intent layer” is the only part of your stack that will appreciate in value.
Keep Logic Thin and Replaceable
Avoid deep, nested dependencies on specific agent frameworks. If a piece of logic can be handled by a better prompt or a more capable model, let it be. Every line of procedural code you write today is a candidate for deletion tomorrow.
Focus on the Data, Not the Wrapper
The “how” of RAG (Retrieval-Augmented Generation) or tool-use is becoming a commodity. The “what”—the proprietary data and the unique insights your organisation possesses—is the actual moat. Ensure your architecture focuses on the delivery of high-quality context rather than the complexity of the retrieval mechanism itself.
The winners in this space won’t be those who built the most intricate scaffolds. They will be the teams who understood that the scaffolding was always temporary, and who focused on the one thing that doesn’t disappear: the clear, calibrated communication of human intent to a machine that is finally ready to listen.
Footnotes
- Rich Sutton, “The Bitter Lesson” (2019). http://www.incompleteideas.net/IncIdeas/BitterLesson.html
- Daniel Miessler, “The Two Types of AI Scaffolding” (2024). https://danielmiessler.com/p/two-types-ai-scaffolding/
- Matthew Berman, “AI Models and the Death of the Wrapper” (2024). https://www.youtube.com/@matthew_berman