The AI Sandwich: Where Humans Excel in an AI World

Introduction to Compon Engineering

Compon Engineering is a philosophy and framework initially developed for engineering workflows, but its principles apply broadly to product design, knowledge work, and other creative fields. Created to optimize how humans use AI to do better work more quickly, the framework reconstructs the traditional workflow to integrate AI agents effectively.

The Original Four-Step Framework

The initial iteration of Compon Engineering relies on four primary phases:

• Planning: Creating a comprehensive and clear plan so the AI agent knows exactly what needs to be built or done.
• Doing Work: The execution phase where the AI agent implements the plan (e.g., writing code, generating designs, or completing the specified tasks).
• Review: Evaluating the output produced by the AI. This is similar to a traditional code review or peer review to identify areas for improvement.
• Compound (The Most Important Step): Any learnings, corrections, or insights discovered during the planning or review phases are “compounded” back into the system. This knowledge is stored within the repository so that the AI agent can reference it in the future, ensuring it does not repeat past mistakes.

The Evolving Role of the Human

As AI models have advanced, the traditional phases of work have shifted in their reliance on human input.

• The “Work” Phase is Solved: If given a good plan, Large Language Models (LLMs) are currently highly capable of executing tasks, following steps, and doing deep work for extended periods. The manual execution phase is largely automated.
• Planning and Review are Improving: AI is becoming increasingly proficient at assisting with both the creation of plans and the initial review of outputs.

With the middle execution phases automated, the critical question becomes: Where does the human actually fit into the workflow?

The “AI Sandwich” Mental Model

The relationship between human knowledge workers and AI can be conceptualized as an “AI Sandwich.”

In this model, humans are the “bread” on the outside (the beginning and the end of the process), while the AI is the “filling” in the middle (the execution).

Top of the Sandwich: Ideation and Brainstorming

Before a concrete plan can be made, heavy human involvement is required at the very beginning of the product or problem-solving cycle. This involves:

• Ideation: Going wide to generate diverse concepts, angles, and possibilities.
• Brainstorming: Working through poorly understood problems by asking deep questions and exploring the core issues.
• Human Requirement: The human must think deeply and stay actively in the loop to direct the AI. Once a solid brainstorm is complete, the AI can then take over to create the actual plan without human intervention.

The Middle: Automated Execution

The middle of the sandwich is rote work. Once the problem is defined and the plan is set, the AI takes over to execute the steps, write the code, or generate the initial drafts.

Bottom of the Sandwich: Taste and Polish

Once the AI produces an output—even if it technically meets all automated testing requirements and specifications—a human must return to the loop at the very end.

• The “Feel” Test: A human must look at the output, interact with it, and evaluate how it feels.
• Elevation: This is the phase where the work is polished, made beautiful, and elevated from generic output to something uniquely high-quality.
• Taste: If humans do not apply their unique taste and final polish at this stage, all AI-generated work will trend toward generic homogenization. The bar for quality is constantly rising, making human refinement essential.

Why AI Cannot Replace the Human Element

While AI will eventually be able to simulate user personas to aid in brainstorming, there are deep, structural reasons why it cannot entirely replace the human at the edges of the sandwich.

1. Setting the Frame of the Problem

Humans possess the unique ability to shift the “frame” of a problem.

• Example: If a user complains that their knee hurts, an AI within a narrow frame might successfully automate acquiring pain medication (like Advil).
• However, a human can zoom out and change the frame entirely, realizing the actual solution is to stop running on hard surfaces or to stretch a specific muscle.
• AI struggles to independently identify when a frame needs to be changed; the human’s job is to set the bounds within which the AI solves the problem.

2. Lack of World Context and “Taste”

LLMs can be thought of as isolated super-intelligences disconnected from the real, evolving world. Because their outputs naturally default to generic averages, they require heavy human guidance to tune solutions to highly specific, nuanced contexts. Rare expertise and unique personal worldview are required to create art or products that genuinely resonate with other humans.

3. The Reality of AGI Timelines

Artificial General Intelligence (AGI) is often defined as the point where it becomes economically viable to run an autonomous agent 24/7 without a schedule or heartbeat. Currently, AI cannot autonomously finish a task, independently decide on the scope and timeline of the next abstract task, and run continuously without human prompting. Fundamental architectural changes to how LLMs learn are required before this is possible.

Work as Art and Performance

All work exists on a spectrum from rote execution to creative art. As AI automates the rote tasks in the middle of the spectrum, humans are freed to focus on the ends of the spectrum, which behave much more like art or performance.

• The Musical Metaphor: An AI music generator can create a structurally sound song, but it cannot replicate the intention of a composer creating something from nothing, nor can it replicate the unique, expressive interpretation of a live human performance.
• Applying this to Knowledge Work: The beginning of a project (conceiving the idea) and the end of the project (tuning the final product so it “feels” right) are akin to composition and live performance. Everything in the middle is just practice and rote work, which is perfectly suited for AI.

Adapting to the Future of Work

The widespread integration of AI agents does not mean the end of human jobs, but it does mandate a shift in how humans approach their roles. Software engineers, writers, and designers will transition to acting more like managers and product strategists.

Key Takeaways for Future-Proofing Skills:

1. Do Not Cling to the Middle: If your sole value is writing standard code or doing rote execution, you must evolve.
2. Lean into Joy and Taste: Identify what parts of your work bring you energy and joy—whether that is crafting beautiful UI, designing elegant software architecture, or writing resonant copy.
3. Use AI as Leverage: Offload the draining, repetitive parts of your job to LLMs so you have the time and energy to focus obsessively on the beginning (strategy) and the end (taste and polish) of the process.

Meta

Added: 2026-04-25