Core ideas:

• If a feature makes agentic work better, the Big AI Companies will add and maintain the best version
• Bespoke harness strategies are only useful for bespoke SDLC needs, or for a few months of uncopied novelty
• We developers need to pragmatically discern where to focus our creative energies

This is about a John Henry story happening in layers - developer skill vs machines, indie devs vs giant corporations. There is an element of mourning as skills are recategorized from professional to hobbyist, and as every advantage John Henry creates becomes a roadmap item for the machines to self-embed.

Developers are irresistibly drawn to creating tools for their tools. I see endless tutorials and iterations on setting up workspaces, knowledge graphs, multi-agent management, prompt caching, etc. But if you’re working in an agentic environment like Cursor, Codex, Claude Code or Antigravity, those companies are far more incentivized and equipped to embed those features. Ironically, some of the projects I see are actually lagging behind the Big Companies, because developers don’t realize how complex those enterprise products are. A billion-dollar AI company can copy the best indie ideas in their sleep, literally.

All the obvious features started as tinkerer projects, and immediately became table stakes in the industry - prompt caching, file indexing, model selection, git-based versioning. Advanced features like custom LLMs for file compare and knowledge graphs, or whatever the latest white-paper open-source improvements are, have mere days of Twitter fanfare and blog posts before the Big Companies embed an elite version.

As researchers and builders, we still need to keep “sharpening the axe”, understanding the tech and dreaming about better systems. But increasingly, we need to be aware of what the tools are already doing for us, and when they’re good enough, and humbly recognize that shareholder-value-driven companies with agents have maximum incentive. An esoteric, fad harness can be just as confounding to maintain as vibe-coded slop.

In 2024, when Copilot demonstrated somewhat serviceable coding capabilities, many indie tinkerers like myself built git-enhanced loops around the LLMs to facilitate fully vibe-coded work. Those who were not in R&D were astounded by the demo of a prompted static web game. To us then, and to everyone in retrospect, the latest developer tools were inevitable. It’s the same thing now. We can spend our cycles doing free R&D for the Big Companies (and because it’s fun and irresistible to tinker, I will!), but we also have to start categorizing work into “uniquely mine to figure out” and “inevitably Sherlocked”.

AI = LLM + agent capabilities + human interface.

Core ideas:

• Our paradigm for AI assistance needs to be “new coworker”, not “out-of-the-box help”
• Our primary responsibility is providing maximum context
• We, and the AIs, need to introspect the meta of our work. “To what end?” and throughout everything, “why”, even if the answer is unknown.

In 2026, most people are adapting to AI by supercharging work, doing tasks faster and with fewer breaks, rather than revisiting the purpose of the work and whether the ends could be accomplished in entirely new fashion. For software engineering, this looks like a bizarre, ironic performance of agents acting out processes that were built around human engines - my agent writes a work ticket, then writes and pushes code to a branch; your agent writes comments for my agent to respond to, my agent updates the code and the ticket. Other fields are engaging in a similar speed-up of tedious tasks - writing documents, summarizing meetings, doing research. For a variety of reasons, we are building “faster horses”. It is incredibly difficult to challenge familiar, obsolete ceremonies that have transcended practicality and become cultural norms. We have accidentally idolized the meta.

This is the time to be in “discovery” mode around process, fixating on goals and imagining how these synthetic minds might drive solutions. To dare to ask “why” of every engagement. Why do we have this meeting? Do we need a dozen different people for a dozen different disciplines? What if we got rid of email? Or better yet, inviting the synthetic minds into the brainstorming process. There is not yet a set of first-principles in working with increasingly capable AIs.

There are two enormous learnings from the past couple years about the efficacy of agentic systems:

1. AI does better with more context
2. AI is increasingly good at higher-level problem solving; i.e. it can distill bigger-picture problems into tasks rather than just completing tasks

The first is unsurprising, but will be a ubiquitous source of frustration forever. “No, not like that”, “Stop fixating on that one document that I know is obsolete”, “Why do you keep making this mistake?”.

The second is obvious to practitioners in disciplines like mathematics and software security, who have seen it first-hand, but met with skepticism by culture at large. It will take inevitable advances in AI comprehension to demonstrate the same proficiency in disciplines like plumbing and cooking.

My proposals for 2026-2027 basic principals in working with AI are these:

1. The paradigm is the same as onboarding a new coworker. They have general knowledge, they’re dumb at some things and good at others, they’ll make mistakes, they’ll ask too many or too few questions. If they’re consistently inept, fire them, but we need to expect training periods.

2. Context is the engine’s steering wheel. Our job is not to bolt-on AI to work faster or harder, but to fill it with context. Documents, meetings, institutional knowledge, casual thoughts, introductions. The AIs will have surprisingly good ideas about how they can contribute after they have lots of context. Context windows will grow.

3. Keep involving AI in the meta. Ends, rather than means. The goal isn’t to fill a spreadsheet, it’s to understand the inventory; maybe there’s a better solution than spreadsheets. Introduce your coworker to yourself and the company, and the top-level vision. It’s more work for the expert human to start with the highest context and work their way down, but only up-front. It’s much more work to keep correcting the coworker with previously hidden knowledge. Help the AI to see the forest for the trees, and to get good at understanding related, cross-disciplinary work. By inviting it into the meta, you’re giving it a promotional track, foundational expectations, opportunities to suggest big-picture ideas. This helps us humans refine the meta, too.

Unfortunately, using agents this way is tough unless you’re an engineer with a good budget and a lot of freedom. Tools like Claude, Codex, and Cursor are unapproachable for late-adopters, operating systems are inexplicably devoid of frontier-model capabilities, and ecosystems like Apple and Google seem to have no unified vision.

While this post can stand alone, it builds on Part 1.

When our faces get too hot, we can vainly wave our hands to push air at them. It’s convenient when a fan replaces human-powered cooling; it’s even better when spinning fans advance beyond human ability. But it is transformative when a fan is strong and fast enough to be a helicopter.

Abilities that lie beyond some barrier of velocity are Speed-Locked Enablements.

Usually, a law of physics, an effect of human perception, or a combination thereof, brings about the unlock. A fast enough rhythm is a musical note. A rapid sanding system becomes a firestarter. A succession of pictures renders an animation.

Initially, scaling speed just scales the task. A shovel that is a thousand times stronger and faster, like a backhoe, just digs bigger and faster holes.

Consider the evolution of calculators, starting with the abacus. Let’s say an abacus can do five arithmetic operations per second. It’s a serious gain for most number-crunchers, but ultimately, they’re simply crunching numbers faster, hitting quotas or leaving work early. It’s an enhancement as opposed to an enablement. The first mechanical calculators were actually slower, but they teed up enablement speeds by replacing the human engine. Early computers, including electronic systems like the ENIAC of the 1940s, were essentially (albeit reductively) conveniences. They were time-savers and accuracy enforcers, and made tedious math problems, like flight trajectories, more palatable.

In 1951, MIT’s Whirlwind I computer broke the speed barrier to an enablement: it could simulate an aircraft’s trajectory in real-time. The transformation was not a result of a new core mechanism - all computing devices (excluding quantum) are just simple arithmetic operations on countable objects, be they marbles or charges. The transformation was the result of speed.

When the crunched numbers represent pixel colors, and the math is fast enough, we get Graphical User Interfaces. Go faster, and we can navigate spacecraft, simulate crashes, or render the sun reflecting off an orc’s swinging axe. All the parallelization, optimization, and specialized components (like GPUs) that enable digital experiences promote on the same simple arithmetic as the abacus; they just spin the math-wheel faster. Phones can do billions of bitwise operations per second, beefy computers can do trillions.

Which takes us to the LLM, the abacus of artificial reasoning. Like the abacus, it started at human-ish speeds, with words-per-second instead of calculations (“words” being a more friendly unit than “tokens”). We’d type or say some input and watch the AI write back at similar speeds, maybe 5-10 words per second - again, like the abacus. With initial speedups, we got faster chatbots. But what happens at a million words per second? A billion?

Can hallucinations be drowned out by instant quorum? What kinds of architectures will be built to parallelize, orchestrate cores, augment LLMs like the CPU? What abilities unlock when an AI has a trillion collaborative ideas per second, or a robot can ponder, “What should I do next?” after every micrometer movement and change in the wind? What if we could simulate the collective individual responses to a policy change or price hike or opening of park? What if technology was so responsive that it was designed in the instant at hand? What if each human’s need was like a pixel, and the AI was fast enough to solve them all in a way that formed a holistic picture?

GPT made this timeline of CPU speed and abilities

The Hypothetical Museum, Pt II (human-authored)

<- Part 1. The launch of the Era of Synthetic Reasoning had been fittingly momentous, by all accounts. And there were many accounts, as every imaginary soul in the Hypothetical Museum of Human Activity had gathered outside the brand new Room of Reasoning to witness the occasion. Old ceremonial texts had to be dusted off to do the thing properly; it had been a long, long time since a raw human ability like “digging” or “travel” had received its first tool. Many there had never considered that “reasoning” might exist, with any utility, outside an organic mind. That irony had been playfully worked into the presentation - books were borrowed from the Library of Everything and select passages read that more or less declared “the pure humanness of reason.”

The crowning moment was the Anointing of the Prime Ancestor - placing an artifact on the pedestal that marked the beginning of a new timeline of inventions - this first “Tool for Reasoning”. The Large Language Model had been chosen, represented by a sort of Platonic ideal LLM with just enough GPT-2 vibes to qualify as a corporeal tool. Of course, the choice of LLM was preceded and followed by much deliberation. Why not the more general concept of “AI”? Or the Transformer Architecture that made it work, or even the Deep Neural Networks that were surely the porting mechanism from biological reasoning to the synthetic version. But debates were always part of such occasions; seasoned curators remembered the tense conversations about which sharpened stick or rock counted as the “First Digging Tool”. A placard attempted to capture all the concepts, generally concluding that AI-related technologies had brought about an auspicious new thing to be endlessly used and improved. Further nitpicking was left to the Museum’s historians and philosophers.

Everyone else soon turned their attention to the new timeline itself, watching for the inevitable emergence of descendent creations. They were not disappointed. Daily, often multiple times per day, couriers rushed in models with multi-modal inputs, novel training mechanisms, attention-head modifications, parallelization, dedicated chips, even diffusion-based language models. The growth was so rapid that the Museum Architecture Team had to move up expansion plans by decades. (They defended original estimates by pointing out that, after Stairs and The Pulley were invented, the Methods of Human Ascent exhibit had served, almost unaltered, for centuries.)

Even more staggering was the proliferation of Supported Inventions. A nicety of the Museum’s hypothetical nature is that fanciful features are trivial to implement. Features like Support Tracers, glowing lines that emanate from artifacts, connecting to tangential, but dependent, inventions outside their strict lineage. For instance, the Internal Combustion Engine radiates many Support Tracers, one of which meanders through the Museum for a half kilometer before arriving at the 1997 Honda Civic.

Like the combustion engine, the LLM was soon tangled in a dense web of glowing Tracers, as startups and research groups rapidly built atop it. To better estimate growth rates, the Architecture Team looked for recent inventions with similar trajectories. The CPU was the obvious choice, already brightly pointing to the LLM with a thick Support Tracer. But unlike the LLM, the CPU itself was not a First Version; it was a clear descendant of the First Enhancers of Human Calculation, like the abacus.

The abacus was brilliant and useful and helped change industries, but it was not renowned as a catalyst for other inventions, as the LLM and CPU were. In some parts of the Non-Hypothetical World, it was still in use after more that 4,500 years. Like most First Enhancements, the abacus made work moderately easier, though there were still elite humans who could beat the machine-users. What made the Museum’s Architecture Team extremely nervous was the steady, predictable set of Support Tracers from the abacus family - accounting systems, currency conversion, land surveys - right up until the early 1950s. There, the massive, five-thousand vacuum tube Whirlwind I sat, with a disconcerting Support Tracer labeled “REALTIME defense systems”.

See also:

https://en.wikipedia.org/wiki/Beta_movement

Computing history

https://en.wikipedia.org/wiki/Timeline_of_computing_hardware_before_1950

https://en.wikipedia.org/wiki/History_of_general-purpose_CPUs

https://en.wikipedia.org/wiki/Mechanical_computer

https://en.wikipedia.org/wiki/History_of_computing_hardware

https://en.wikipedia.org/wiki/Abacus

In the Hypothetical Museum of Human Activity is a display for every artifact of humanity - every work of art, every object of worship, every toy and tool. Hypothetical archeologists and anthropologists have united to painstakingly unearth, restore, and classify pieces of pottery, carved idols, cave paintings, etc.

Within the Hall of Farming is the Evolution of Plows exhibit, a branching timeline of simple sharpened sticks and stones, progressing, several vaguely parallel inventions at a time, to the John Deere 2730 Combination Ripper. Down the hall stands the entrance to the Palace of Containers, home to an even more complex family tree of mud pots and pyrex dishes. The Museum’s East Wing hosts the Chamber of Computing. Tools here are far younger; a mere 5,000-year-old abacus rests atop a pillar, beside its cousins. These abaci are the Sharpened Sticks of Computing, the First Enhancers of human calculation. They are followed, with enormous lulls in the timeline, by a mess of mechanical labors of love and, eventually, a variety of metal and silicon cuboids of decreasing size. This display has been a recent hotspot for imaginary visitors, but today it is empty.

Today, the could-be crowds have gathered, with almost as much excitement as the Museum’s caretakers, to witness the opening of a new room. They will watch in reverent wonder as a lone artifact is placed on the ceremonial pedestal that marks the beginning of a new timeline. And not just any timeline. They were there to celebrate the opening of Human Flight and debate whether The Airplane actually warranted a new timeline or was merely a branch off Vehicles. They were there to nod in begrudging agreement when the first dedicated Back Scratcher Stick took its place on a pedestal. Today’s timeline is different, not as flashy as the light bulb or immediately miraculous as vaccines, but in some ways more prestigious. It’s a class of timeline that has not been kicked off for millenia.

For all our glaring lack of tails, scales, wings, and claws, humans are still equipped for a huge variety of tasks, even without tools. On the mental side, there’s intricate communication, planning, and a world of creative art and engineering. Physically, our headlines tend to be “opposable thumbs” and “able to travel for many miles”. But we’re also Swiss-army-knives of less lauded skills:

• hold water by cupping hands
• exhale a targeted breath to repel an insect
• fingernails to remove splinters or pry open nuts
• spit to moisten
• strategic joints and bones to break sticks
• sundry other “party tricks of survival”

Of course, it’s tool crafting and usage that sets us apart. For this paradigm, we’ll group tools into two reductionist categories:

1. Enhancements - improve existing human capabilities. A drinking glass or bowl can hold water better than our hands.
2. Enablements - facilitate entirely new capabilities. Humans can’t emit light, but we’ve invented luminous tools.

For both Enhancements and Enablements, “Version 1” is momentous for our species. The first cave-brightening, fire-on-a-stick torch. The first ground-tilling sharpened rock. That spreading, cultural moment of, “Wow, we’ve only ever done this bare-handed; now there’s a way to go beyond our limits.”

Nearly every raw human capability has long since been entrusted to iterations of tools. Travel, carrying possessions, creating fabric, cleaning floors, planting, watering, harvesting. Most new Version 1s are Enablements. As far as Enhancements go, what Version 1s are even left? What can a human do that hasn’t been addressed by invention?

That’s the incredible thing about AI - unlike every Version 1 invention in our lifetimes, AI is not an Enablement. For the first time in millennia, a new Version 1 Enhancement has emerged; it does something humans already do. Even crazier, AI has produced multiple Version 1s within a few years. Wilder still:

Every remaining Version 1 Enhancement, i.e. every objective thing that humans can do, now has a tool.

For better or worse, AI can make music, draw pictures, write stories, converse, and reason. Like all Version 1s, these tools will be improved. Version 1s are not incredible because they’re the best, or even better than skilled humans; they’re incredible because they herald an end to working within the limits of a raw human capability. As we consider analogous technologies to predict AI’s impact, “The Internet”, “Computers”, and “Engines” may fall short. A more fitting comparison of anthropological impact would be Stones For Counting or Wheels For Travel.

These Version 1 Enhancements broadly represent tools for “Creating Art” and “Reasoning”. Evaluating art takes us into the philosophic, as does the question of whether AI (specifically LLMs) can actually reason like a human. Practically, the takeaways are that we have Version 1 Robot Artist, and Version 1 Reasoning that are good enough for some practical application. The questions of, “what does this mean?”, “what can we do now?”, “what should we do now?” are perhaps more critical than ever, but also standard for any invention.

What we’re uniquely left with, now that every aspect of work as it relates to industry and chores can be mimicked, are those eternal philosophers’ questions:

“What is it that humans uniquely do? Love? Critique? Be silly with friends? What does it mean to bring simulation to those?” “Are humans more than physical? Are spiritual experiences a physically stimulated illusion? Does it matter?”

“What is purpose? Legacy? Why are we here?”

What’s more, the very creation of LLMs is based on a scientific hunch that a brain-like architecture, with brain-like training, might perform brain-like tasks. There is identity crisis in the entangled unknowns of how our minds work and whether these creative models themselves might hold insights or even romance-shattering answers.

Certainly, AI will spur and spawn untold inventions, including new Enablements; like the Wheel before it, we can’t help but imagine the descendent line of carts, cars, and Mars Rovers that a Version 1 inspires. We’ll have an endless supply of “back in my day” and man-vs-machine rifts that are as old as invention.

But I anticipate that AI will also bring about a resurgence of focus on human meaning, both demanding and enabling a convergence of philosophic, religious, and moral thought about the Grand Meta. And unlike past events that forced non-philosophers to answer the Big Questions (usually times of fear and uncertainty), this one will not pass. Behind the commotion of new progress and invention will be the shrinking list of uniquely human activities, which could fast become the only non-commodities.

Part 2 ->

See also:

Interview with Ilya Sutskever, the father of modern LLMs, conducted wonderfully by Jensen Huang: https://www.youtube.com/watch?v=I6qQinoY9WM

Anthropology of Tech: https://en.wikipedia.org/wiki/Anthropology_of_technology

Some very old, very controversial thoughts on the matter: https://www.bible.com/bible/111/ECC.1.NIV