The AI Sandwich: Where Humans Excel in an AI World

The AI Sandwich Metaphor and Compound Engineering Framework

The "AI Sandwich" metaphor describes a work approach where humans serve as the "bread," framing problems at the beginning and reviewing outputs at the end, while AI acts as the "filling," handling the core, automatable tasks. This model suggests that full automation often misses the mark when personal or creative input is essential.

This framework highlights that to maintain quality and ensure work feels like "your own," human involvement at the bookends of a process is vital, especially as output quality expectations rise.

Kieran Klaassen, GM of Korra, is introduced as the creator of Compound Engineering. This engineering framework and plugin is widely adopted by those coding with AI agents, including internally at Ever.

Klaassen, along with Trevin Chao, developed Compound Engineering, which is seen as having major implications for the future of programming and the overall impact of AI on various work processes.

Humans are the bread in the sandwich, and the AI is in the middle.

Compound Engineering: A Four-Phase Workflow for AI Agents

Compound Engineering is a methodology initially developed for AI-assisted engineering work at Cora, but it applies broadly to product, design, and general knowledge work. It outlines a structured process for effectively collaborating with AI agents to improve work quality and speed.

The workflow begins with "Planning," where a clear understanding of the task is established. Next is the "Work" phase, where an AI agent implements the plan, whether it involves writing code, designing, or other tasks.

Following work, the "Review" phase assesses the output for quality and areas for improvement, similar to traditional code reviews. This step identifies what went well and what needs iteration.

The "Compound" phase is critical: any insights or lessons learned during planning or review that could prevent future errors or improve processes are integrated back into the system, ensuring continuous learning and improvement for the AI.

The most important step is the compound step, which is if anything comes up during that review or during the planning that you think like, 'Oh, this is, this is good learning. Probably you will run into this again. You can compound that knowledge back into the system.'

LLMs efficiently automate the 'work' phase in agentic coding

In agentic coding, the 'work phase' – where an agent executes a predefined plan – is now largely considered a solved problem. Large language models have proven highly effective at performing this stage, demonstrating a capability to follow complex steps and carry out deep work for extended periods, sometimes hours or even days.

This efficiency means that once a robust plan is in place, the model can be trusted to execute it with minimal human oversight. The need for constant human intervention during the task execution is significantly reduced, allowing developers to rely on the LLM to follow through on the planned actions.

The effectiveness of LLMs in the work phase shifts the focus for human developers. While planning and review phases are also seeing improvements with AI, the bottleneck for human involvement is no longer primarily in the execution itself, freeing up time for more strategic or creative tasks.

LLMs are very good at just following steps, doing deep work, like working for hours, days even now, and that, that thing is kind of solved.

Humans Define and Elevate Work in an AI Sandwich Workflow

In AI-driven workflows, humans are crucial at the start and end, forming an 'AI sandwich' model. The initial phase involves deep human engagement for brainstorming, ideation, and problem framing, where individuals ask many questions and think hard, while the LLM supports. This ensures a strong foundation for the work.

Once the problem is well-understood, the middle phase is largely automated. AI can efficiently create a detailed plan, eliminating the need for constant human oversight during this execution-focused stage. This frees up human cognitive resources for more complex tasks.

The final stage requires human involvement again to polish and elevate the output. Even after automated testing confirms functionality, a human's critical eye can identify areas for improvement, enhance design, and make the work 'feel good.' This dedication to refinement prevents generic, undifferentiated results.

This sandwich approach highlights the importance of knowing when to engage human thought versus offloading tasks to algorithms. By focusing human effort on high-value creative and quality control steps, the overall output is elevated and meets a continually rising bar for excellence.

The beginning and the end, and the middle is kind of soft and can be automated pretty well, and... that's now referring to the AI sandwich.

Humans excel at setting and shifting problem frames

AI agents are increasingly capable of handling tasks like ideation and brainstorming within a given problem. However, a critical distinction remains in how problems are defined, rather than just solved. While agents might progress in automating processes, the fundamental human advantage lies in the initial conceptualization of the challenge.

Consider a problem like "my knee hurts." Within a local frame, the immediate solution might be to "take Advil," and many steps, such as getting the medication, can be automated. This represents solving a problem within clearly defined, immediate boundaries, where the goal is a direct, surface-level fix.

However, humans have the unique capacity to see the "larger frame" of the same problem. Instead of just taking Advil, the knee pain might stem from an unaddressed issue like a tight IT band or running on hard surfaces. This involves redefining the problem at a different level of abstraction, moving beyond symptoms to root causes.

This ability to "flip and change frames" is a core human skill. It means humans can continuously redefine the scope and context of a problem, setting the boundaries for what needs to be solved. This ensures that even as AI becomes more capable, humans remain essential in determining the fundamental questions and structures that guide problem-solving.

Humans are very good at flipping and changing frames like that, and our job is to set the frame, set the bounds within which we solve.

Human Taste and Experience Are Essential Amidst Automation

Humans with experience and taste are crucial for discerning quality, deciding if something is "shit or great." This judgment allows individuals to claim ownership over their work by making the final "yes" or "no" decisions, which cannot be fully automated if the goal is personal expression or ownership.

While AI can simulate responses from millions of people, the target of full automation is constantly moving. New layers of complexity and higher-level decisions emerge, always providing opportunities for human expertise to make a significant impact.

Rare expertise, cultivated over a career and embodying a personal worldview, is challenging to capture in language models due to its scarcity. This unique human insight provides critical feedback loops that AI struggles to replicate, ensuring that human judgment remains irreplaceable.

The effective strategy for humans in an increasingly automated world is to "ride the model" or act as the "bread in the sandwich," leveraging AI's capabilities while infusing the work with unique human judgment and experience.

The simple answer is, ride the model. or to mix the metaphor, be this, be this bread in the sandwich.

Engineers Will Focus on Product and Beautiful Output in an AI World

In an AI-accelerated future, the nature of work for professionals, particularly engineers, will evolve significantly. Simply writing code will no longer be enough; the focus will shift towards creating exceptional, or "beautiful," output across various domains.

This shift implies that engineers will become more product-focused, taking on responsibilities akin to product managers or product engineers. Instead of solely focusing on functional code, they will be encouraged to pursue excellence in aspects like beautiful code, elegant abstractions, refined architecture, compelling design, or impactful copy.

The core idea is to identify and lean into the aspects of work that genuinely bring joy and energy. By focusing on creating "beautiful" outcomes in their chosen domain, individuals can leverage AI tools, such as Large Language Models, to enhance their work rather than feeling drained by it, transforming their professional output into a source of energy.

lean into making beautiful stuff and whatever that means To you, that can mean beautiful code, beautiful abstractions, beautiful architecture, beautiful design, beautiful copy, I, I think it's very important to lean into what is beautiful to you because Then you will find a way to utilize an LLM to make something that gives you energy instead of drains you all the way.

Defining AGI as a 24/7 Autonomous Agent

Dan Shipper proposes that Artificial General Intelligence (AGI) should be defined by its economic profitability to run as a 24/7, always-on agent. While current models like OpenClaw are pushing towards this, they still operate on schedules and heartbeats, unlike an agent that continuously works and spends tokens because its output is consistently worthwhile.

A key limitation of current models is their inability to autonomously change frames or select new tasks after completing a specified one. Although they can execute well-defined tasks for extended periods, they do not independently transition to the next objective. Achieving true 24/7 autonomous operation will likely require fundamental changes to language model architecture.

This extends to creative fields, as Kieran Klaassen points out using music as an example. While AI can create songs, it lacks the internal human experience and context that defines a live performance or the act of composing a melody. The unique human element in creating and delivering art, which evokes a deeper feeling in an audience, is still missing from AI-generated outputs.

When models can run constantly, they would likely be sensitive enough to context to perform truly creative tasks, bridging the gap between generic output and meaningful, human-like creation.

My general bar for AGI is whenever it is economically profitable, or makes economic sense to run an agent 24/7, or it never turns off.

Art as a Metaphor for AI Collaboration: Start, Rote, Performance

Kieran Klaassen draws a parallel between the structure of an AI collaboration and the process of musical composition. He identifies the initial spark of "coming up with a piece" as a special, creative moment, akin to starting a new project with AI.

The middle phase of a musician's work, which involves repetitively practicing a piece a hundred times, is described as rote and less creative. This is the part of a project that AI is well-suited to handle, taking on the repetitive, mechanical tasks.

The final stage, the "performance" or the "Polish step," is another special moment where the completed work is presented to others. By automating the rote tasks with AI, humans can dedicate their energy to these more impactful and creative stages of ideation and refinement.

This perspective suggests that leveraging AI allows individuals to shift their focus from mechanical work to the more artistic and interesting aspects of creation, such as generating novel ideas and perfecting the final output for an audience.

Practicing a piece for like playing it a hundred times is not very creative as a musician, and this is kind of the middle part. But at the end, the performance is where you bring it out into the world, to the people. So I think that's a special moment.

Embrace creative work that brings joy in an AI-driven world

The concept of what constitutes "rote" work versus "art" is constantly shifting. As AI automates routine tasks, humans are pushed towards more creative and expressive forms of work. This dynamic means the definition of essential human contribution continues to evolve.

The ultimate non-automatable work is art created by humans driven by genuine feeling. While this might seem abstract, everyone possesses a unique creative spark. It is about identifying what truly brings joy and excitement.

Individuals should actively pursue and lean into activities that light a fire in them, whether that is beautiful writing, structured lists, or any other source of happiness. Integrating AI, such as LLMs, into these joyful tasks can enhance and empower this creative pursuit, making work more fulfilling.

The final thing that's not automatable is, is like art made by humans who feel something, and I think that's beautiful.