# 22 Truths About Where Software Is Headed *Mathieu Glaude — Last updated: March 23, 2026* These are the truths I keep coming back to: where the software industry is headed, what survives, and what doesn't. They serve as a reference for deciding where to invest time, energy, and money — a working thesis for a world where AI changes everything about how software gets built, sold, and consumed. This is a living document. Truths will sharpen, some will be proven wrong, and new ones will emerge. Each update incorporates what I'm learning from building with agents daily and from the practitioners and leaders shaping the frontier. --- # Part I: The Industry Shift What's happening to software, platforms, and the businesses built on them. --- ## 1. Every product becomes an API Whether you like it or not, your software product is becoming an API. AI agents can reach into any system: your CRM, your platform, your toolchain. When an agent can do what a user used to do, the switching cost between platforms drops to near zero. Platform lock-in is eroding. The stickiness that SaaS companies relied on for decades is weakening, not because the products stopped working, but because agents make switching costs approach zero and building custom alternatives becomes viable. Every app with a web interface is now a "slow API" that agents can access through browser automation. Companies that proactively offer agent-friendly APIs will thrive. Those that resist will simply get scraped more slowly. > "Every app is just a very slow API now, if they want or not." — Peter Steinberger, OpenClaw (Feb 2026) Companies that try to block agent access (CAPTCHAs, bot detection, API restrictions) will lose users to agent-friendly alternatives. Steinberger watches his agent "happily click the I'm not a robot button." If you push back too long, you become Blockbuster. > "These apps that are in the app store for using these smart home devices... shouldn't even exist. Shouldn't it just be APIs and shouldn't agents be just using it directly." — Andrej Karpathy, No Priors (March 2026) But APIs alone aren't enough. The real product surface is the "agent harness," the documentation, rules, skills, and domain context that let an agent extract maximum value from your system (see [Truth #7](#truth-7)). ## 2. Context is the product, not software You don't need the platform. The platform often becomes a limiting factor. What you need is the context, structured, well-articulated, in the right format. The future of consuming any service is: "This is the format we need your context in, and that's all you need to give us." Every business that can't articulate its value as context will be replaced by one that can. The agent's power comes from having rich context about the user, their situation, their history, and their preferences. Data and context beat features. The company that holds the most structured, relevant context wins. There's a deeper reason context wins: AI is fundamentally a compression technology. It takes the entire internet, all human knowledge, and compresses it into a few hundred gigabytes. Most enterprise data is massively redundant; the same information lives in Slack, Salesforce, Google Docs, email, and a dozen other systems. AI naturally compresses away this redundancy. What survives compression is the authoritative, canonical, deduplicated source of truth. Organizations that already operate with canonicalized data (a single schema, a single data dictionary, a single governance registry) are already in the compressed form. They are the Wikipedia of their domain. > "What people underestimate with AI is it's a compression technology... the amount of true novel information and knowledge in human society is quite limited. What we see is repetitions and variations on the same themes over and over." — Sebastian Siemiatkowski, Klarna (Feb 2026) > "It can modify my gym workout based on how well I slept, or if I have stress or not. It has so much more context to make even better decisions than any app even could do." — Peter Steinberger, OpenClaw (Feb 2026) Karpathy's auto-research project makes this concrete: he describes an entire research organization as "a set of markdown files that describe all the roles and how the whole thing connects." The organization's value is its structured context, not the code it produces. The logical endpoint is clear: document repositories die, databases win. All content (policies, communications, strategies, skills, decisions) becomes structured data living in a graph. Markdown and prose survive only for human-facing documentation. Everything agents and systems consume should be structured from birth. The default should be structured data. Prove why it shouldn't be. The most effective way to make AI useful isn't to build general-purpose "AI lawyer" or "AI accountant." It's to narrow the problem domain by encoding your specific business rules as structured context. The more precisely you articulate the constraints, the more you turn an AI science problem into an engineering problem. > "If you narrow the domain — 'here are our core legal terms, here are the axes of independence' — you create a narrow problem domain... you turn science into engineering." — Bret Taylor, Sierra (March 2026) ## 3. Software as we know it is temporary All software does is take data, transform it, and put it somewhere else. That middle layer, the code that does the transformation, is going away for most use cases. It will be replaced by inference. Not today for everything, but the trajectory is clear and accelerating. Code will be one of the first domains where AI achieves superintelligence-level capability. Math and code are falling first. The rest follows. The exceptions are real but narrow: deterministic systems that need provable guarantees, regulated environments requiring auditable execution, and edge cases where inference confidence thresholds aren't acceptable. These don't disappear; they become the interesting part. Most business logic becomes inference; the exceptions are where differentiated value lives. > "In the past, we code. In the future, we're going to write ideas, architectures, specifications. We're going to organize teams. We're going to help them define how to evaluate the definition of good versus bad. What does it look like when something is a great outcome?" — Jensen Huang, All-In Podcast at GTC (March 2026) Question your software all the time. What is its purpose? If the answer is "transform data from one shape to another," that's exactly what inference replaces first. We're already seeing concrete evidence of this at the most AI-native companies. Claude Code, the most widely used AI coding tool on earth, responsible for 4% of all public commits, has no code older than six months. The entire codebase has been written and rewritten continuously as models improve. The creator of Claude Code has uninstalled his IDE entirely and doesn't edit a single line of code by hand. The shelf life of code is now measured in months, not years. > "All of Claude Code has just been written and rewritten and rewritten and rewritten over and over and over. There is no part of Claude Code that was around six months ago. It's just constantly rewritten." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) > "Today coding is practically solved for me and I think it'll be the case for everyone... We're going to start to see the title software engineer go away." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) ## 4. AI commoditizes asset-light businesses For 50 years, the tech industry believed that companies built on software and services, without needing to own physical assets like factories or inventory, were inherently superior. Higher margins, faster growth, less capital required. This was treated as a permanent law. But if AI makes software near-free to create, that advantage disappears, and we revert to the historical mean, where value accrues to physical infrastructure, regulated industries, and proprietary data. > "Every job invented in the 20th Century is threatened by AI." — Nassim Nicholas Taleb (Feb 2026) This is the Lindy Effect applied to labor and business models. SaaS, knowledge brokering, consulting, middle management: these are all 20th-century inventions, and they're fragile precisely because they're new. The things that survive are the things that have always existed: trust, governance, physical infrastructure, regulated accountability. The longer something has survived, the longer it will continue to survive. The 50-year supremacy of asset-light businesses isn't a law; it was an anomaly. What's actually safe when AI commoditizes software: - **Regulated and liability-bearing businesses:** someone has to be on the hook - **Anything touching the physical world:** hardware, manufacturing, energy - **Proprietary data sets:** AI makes your data more valuable, not less - **Marketplaces and network effects:** liquidity is worth more than software - **Operationally intense businesses:** the "bad" businesses become the best ones - **Cybersecurity and physical security:** more AI means more attack surface - **Reusable governance processes:** the operational IP of harmonizing data, establishing evidence, and credentializing information across an ecosystem. Agents can write code; they cannot build institutional credibility or navigate the relationships required to establish trusted data supply chains. - **Distribution and network access:** when products can be copied and open-sourced almost instantly, the channel to users becomes the moat. AI can build the software, but it cannot build the distribution. - **Brand and community:** human trust, reputation, and relationship history compound over time in ways that AI cannot replicate or shortcut. Community creates switching costs that persist even when software costs go to zero. - **Outcome-based pricing models:** companies that price on outcomes rather than seats scale with AI adoption instead of against it. When agents reduce headcount, seat-based revenue shrinks; outcome-based revenue grows with the value delivered. When agents can build custom tooling in minutes, the willingness to pay monthly SaaS subscriptions for generic, horizontal tools gets squeezed. Organizations increasingly see that agents can build custom alternatives for workflows that don't require deep domain expertise. And when billions of coding agents exist, any product can be copied, improved, and open-sourced almost instantly. Defensibility shifts entirely away from code. > "There's a real risk that more people will build than they do now versus buy because the marginal cost of writing software goes to zero. I think that'll be true for some software, particularly developer platforms being consumed by engineers." — Bret Taylor, Sierra (March 2026) > "Five years out, when billions of coding agents exist, software development will be largely solved. The traditional moat of 'we have better engineers' disappears. Products will be copied, improved, and open-sourced almost instantly. Defensibility will shift away from code itself, toward distribution, data, brand, and community." — Kevin Rose (March 2026) > "Why should I pay another subscription for something that the agent can just do now?" — Peter Steinberger, OpenClaw (Feb 2026) The market is already pricing this in, violently. Between January 15 and February 14, 2026, over $2 trillion in software market cap evaporated. Thirty days. Adobe, Microsoft, Salesforce, SAP, ServiceNow, and Oracle alone shed over $730 billion. The S&P 500 Software & Services Index fell 20% year to date while the broader market stayed flat. Atlassian dropped 35%. Salesforce dropped 28%. Their core workflows (task tracking, data entry, customer logging) are exactly what agents automate first. The valuation split tells the story clearly: AI-native companies trade at a median 10x+ revenue multiple. Traditional SaaS sits below 5x. That gap was already wide six months ago. It doubled this year. Meanwhile Palantir, which never sold seats but sold decision infrastructure on top of messy, fragmented data, posted 70% revenue growth and trades at 44x forward sales. The companies that survive are those whose value was never reducible to a per-seat workflow. > "Pure software is rapidly becoming un-investable." — Naval Ravikant (Feb 2026) One important counterpoint: AI doesn't only compress margins; it can expand total capability. Klarna cut headcount by more than half and simultaneously improved their service quality. That's not just cost reduction; it's a step-change in what's possible with fewer people. The outcome isn't always decline; it's bifurcation. Businesses that adapt gain capabilities they never had before. Those that don't face the margin compression in its pure form. The question for every company is which side of that bifurcation they land on. The nuance is that SaaS isn't dying; it's bifurcating. Generic horizontal platforms (task tracking, data entry, CRM workflows) face margin compression because these are exactly what agents automate first. But domain-specific applied AI companies, those that wrap deep business expertise into agent harnesses for specific verticals, represent a massive market expansion. Bret Taylor's position: there aren't enough of these companies yet, and their absence is what's slowing enterprise AI adoption. The threat isn't to all SaaS. It's to SaaS that competes on features rather than domain depth. Jensen Huang offers an important counterpoint to the "enterprise software is dead" narrative: existing tools with deep domain functionality (Photoshop, Synopsis, SQL databases) don't die — they get 100x more users, because agents use them. The software becomes the output interface through which humans verify and control agent work. What dies is generic workflow software. What survives becomes the agent's hands and the human's control plane. > "The enterprise software industry is limited by butts and seats. It's about to get a hundred times more agents banging on those tools." — Jensen Huang, All-In Podcast at GTC (March 2026) The survival mechanism is transformation: enterprise software companies become value-added resellers of AI tokens, bundling domain-specific harnesses with their existing customer relationships and regulatory compliance. Their go-to-market advantage — not their code — is the moat. > "Deep specialization. The call to arms for entrepreneurs is: know your vertical. Know it as deep and as better than everybody else. And then wait for these tools because they're catching up to you and now you can imbue it with your knowledge." — Jensen Huang, All-In Podcast at GTC (March 2026) > "I'm actually extremely bullish on applied AI... if we had a mature applied AI market where the CFO could go buy that agent to onboard new supply chain vendors that just worked, we could accelerate trillions of dollars of economic value. The lack of existence of all those other companies is impeding adoption." — Bret Taylor, Sierra (March 2026) There's a broader historical pattern here: when the cost of producing something collapses, demand rarely stays flat. It expands by orders of magnitude. Personal computer prices fell 99.9% between 1980 and 2015. We didn't consume the same amount of compute more cheaply; we built entirely new industries on top of the cheap layer. AI is now collapsing the cost of cognitive work, and services represent nearly 80% of US GDP. If service costs fall dramatically, the effect functions as an invisible tax cut across the entire economy, and households gain real purchasing power even without wage growth. The companies that die are those charging for what becomes free. The companies that thrive are those building on the newly-cheap layer. > "Even Tailwind... they had to cut off 75% of employees because nobody's going on the website anymore because it's all done by agents." — Peter Steinberger, OpenClaw (Feb 2026) > "I have changed my mind on software. Software is getting much more risky and it's much more unclear what the future of SaaS is." — Sebastian Siemiatkowski, Klarna (Feb 2026) ## 5. Exponential acceleration is real and we can't intuit it Human brains are not good at processing exponentials. From year to year, month to month, the leverage people get from AI and the value they're able to create is accelerating. The shift to the software market and the general economy will be quicker than we think. Today's frontier models become tomorrow's cheap, fast models, within months, not years. The big model you're paying premium for today will be the lightweight model you use for simple tasks next quarter. > "The gap between what AI can do and what most organizations are actually doing with it is enormous, and growing." — Bret Taylor, Sierra (March 2026) The compute scaling alone tells the story. In just two years, AI computation demand increased by a factor of 10,000x across three inflection points: generative (ChatGPT), reasoning (o1/o3), and agentic (Claude Code), each requiring roughly 100x more compute than the last. And we're still early. > "When we went from generative to reasoning, the amount of computation we needed was about a hundred times. When we went from reasoning to agentic, the computation is probably another 100 times. In just two years, computation went up by a factor of 10,000x." — Jensen Huang, All-In Podcast at GTC (March 2026) > "We haven't even started scaling yet. We are absolutely at a million x." — Jensen Huang, All-In Podcast at GTC (March 2026) Small companies with low overhead can deploy capital efficiently and ride the exponential. The cost structure advantage is real and growing. A small team with AI leverage can outproduce organizations 10x or 100x their size. Large incumbents (Atlassian, Salesforce, the big consultancies) can make hay right now, but eventually they're gone. They're trying to stay a week or a month ahead of their clients, flying by the seat of their pants. The numbers are staggering. Since Claude Code launched, productivity per engineer at Anthropic has grown 150%. For context, the creator of Claude Code was previously responsible for code quality at Meta, where a 2% productivity gain required a year of work by hundreds of people. Steve Yegge estimates an Anthropic engineer now averages 1,000x the productivity of a Google engineer at Google's peak. Three years ago we were still talking about 10x engineers. Now we're talking about 1,000x, on top of already-elite baselines. David Friedberg described running autonomous research on genomic data where something that "would normally be a PhD thesis that would take seven years" was done in 30 minutes on a desktop computer — a roughly 120,000x speedup. > "Since Claude Code came out, productivity per engineer at Anthropic has grown 150%... In my old life I was responsible for code quality at Meta. Seeing a gain of something like 2% in productivity was like a year of work by hundreds of people. So this — this is just completely unheard of." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) The implication for management is concrete: token consumption is becoming a leading indicator of organizational AI maturity. Jensen Huang expects a $500K engineer to consume at least $250K in tokens annually. An engineer spending $5K on tokens is the equivalent of a chip designer refusing to use CAD tools. > "That $500,000 engineer, at the end of the year I'm going to ask him how much did you spend in tokens, and that person said $5,000 — I will go ape. If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed. This is no different than one of our chip designers who says, 'Guess what, I'm just going to use paper and pencil.'" — Jensen Huang, All-In Podcast at GTC (March 2026) ## 6. Your real competitor is the AI frontier, not your industry For decades, competitive strategy in knowledge-work industries meant watching your peers, the companies in your sector doing roughly what you do. That framework is now obsolete. IBM didn't post its worst trading day since 2000 because a competing IT services firm outmaneuvered them. Adobe didn't lose 30% of its value to a rival creative suite. CrowdStrike didn't shed $20 billion in market cap because a competing cybersecurity vendor shipped a better product. They all lost to the same thing: a frontier model release that commoditized their core workflow overnight. This pattern will repeat across every knowledge-work industry. The competitive threat is no longer horizontal (peer companies in your sector); it's vertical: the AI frontier itself. A single model release can commoditize an entire category of cognitive work in hours. No traditional competitor has ever moved that fast or that broadly. The strategic implication: stop benchmarking against your industry. Start benchmarking against AI capability timelines. Your planning horizon is not your competitor's product roadmap; it's the next frontier model release. The question is not "what is our competitor building?" but "what will the next model make free?" > "Every technological shift reorders the stack and the companies pricing static workflows WILL struggle. The companies owning data, trust, compute, energy, and verification may thrive." — The Kobeissi Letter (Feb 2026) The companies that survive are those whose value lies in what the model cannot replicate: proprietary data, regulated accountability, physical-world operations, network effects, and trust infrastructure. If your moat is "we do this knowledge work better than the other guys," you no longer have a moat, not because the other guys got better, but because the frontier made the work itself near-free. This competitive dynamic produces extreme winner-take-all outcomes in every market touched by AI. When any category of cognitive work becomes near-free, the gap between best and second-best becomes unbridgeable. There is no demand for average. > "First place gets a Cadillac Eldorado, second place gets a set of steak knives, and third place you're fired." — *Glengarry Glen Ross* (1992), cited by Naval Ravikant on AI and winner-take-all markets (March 2, 2026) The set of niches you can dominate is still infinite. The strategic response isn't to compete broadly. It's to keep redefining your domain until you're the best at something specific. > "Become the best in the world at what you do. Keep redefining what you do until this is true." — Naval Ravikant (March 2, 2026) ## 7. Agent harnesses are the new product surface Products are evolving from two surfaces (UI + API) to three: UI for occasional human use, API for traditional integrations, and Agent Harness for autonomous value extraction. The harness (documentation, rules, skills, guardrails, decision context) is the interface that lets agents extract maximum value from your system. It encompasses the API but adds the domain expertise, decision rules, and operational guidance that APIs alone can't provide. The harness won't be elegant. It's optimized for agent comprehension, not human aesthetics. It may expose every switch in your dashboard, every edge case handler, every decision rule, things you'd never put in a UI because they'd overwhelm a human. But agents thrive on completeness and specificity. The harness is the full instruction manual for your most expert operator, machine-readable. > "The web application of the future... will you have an agent harness? Something more than APIs... the person who's the greatest expert who knows how to extract the most value — that's the harness." — Bret Taylor, Sierra (March 2026) > "There's a lot of switches in the dashboard that there's no API for, and we are all collectively discovering... agents can handle it as long as the harness describes when to use that switch." — Bret Taylor, Sierra (March 2026) The practical implication for every software company: start building your harness now. Document every decision rule, every constraint, every capability your system has, in a structured, machine-readable format that lives alongside your code. The companies that have well-articulated harnesses will be the ones agents choose to integrate with. The ones without will get the "slow API" treatment: browser automation and screen scraping. ## 8. Outcome-based pricing aligns AI incentives Per-seat pricing was the foundational unit economics of SaaS for twenty years. Agents broke the denominator. But the answer isn't just that "SaaS dies." It's that the pricing model must evolve to align with the value agents actually deliver. Outcome-based pricing (charging for resolved cases, achieved goals, delivered value) creates structural alignment between vendor and customer. When the vendor's revenue depends on the agent actually succeeding, they build better harnesses, better guardrails, better domain expertise. The incentive shifts from "maximize seats" to "maximize outcomes." This is the CPC-to-impression-ads shift of AI. > "Outcomes-based pricing... if the AI agent resolves the case, there's a pre-negotiated rate. If we escalate to a person, it's free. This aligns our interests with our clients." — Bret Taylor, Sierra (March 2026) > "When you charge for something closer to business value, it's more efficient. The analogy: impression-based ads to CPC ads. When you're aligned with outcomes, you become a better partner." — Bret Taylor, Sierra (March 2026) The deeper insight: outcome-based pricing also forces vendors to optimize internally. When your margins depend on efficient agent performance rather than seat count, reducing token utilization for the same outcomes becomes your problem, not the customer's. This creates a flywheel: ``` ┌───────────────────┐ │ Better harnesses │ └────────┬──────────┘ ▼ ┌───────────────────┐ │ Better outcomes │ └────────┬──────────┘ ▼ ┌───────────────────┐ │ More revenue │ └────────┬──────────┘ ▼ ┌────────────────────────┐ │ More harness investment│ └────────┬───────────────┘ └───────► (back to top) ``` > "Reducing your token utilization for the same outcomes is your problem, not your customer's." — Bret Taylor, Sierra (March 2026) --- # Part II: The Agentic Future How to build in a world where agents are the primary consumers of software. --- ## 9. Build agentic-first, UI second The majority of the code should be AI. The UI is built on top of what the AI provides, dynamically, in response to what the agent is doing. Don't over-engineer frontends. Don't teach people to use complicated UIs when the AI should be doing the work. Apps are entering terminal decline as a product category. Personal agents have more context, more flexibility, and no subscription fees. People will naturally stop using apps because their agent can do it better. > "Why do I still need an app to do that? Why should I pay another subscription for something that the agent can just do now?" — Peter Steinberger, OpenClaw (Feb 2026) > "I will just naturally stop using [apps] because my agent can just do it better." — Peter Steinberger, OpenClaw (Feb 2026) > "The agent harness will not be elegant because it's optimized for the context needed to perform complex multi-step procedures on behalf of a person. It's just very different than UI design." — Bret Taylor, Sierra (March 2026) Once you're fully agentic end-to-end, you can spin up new features, or entire business offerings, in hours instead of months. ## 10. Build codebases for agents, not for humans The audience for your code is no longer primarily human developers; it is agents. Naming conventions, file structure, documentation, and architecture should optimize for agent navigation, not human aesthetics. Agents start every session from zero. They know nothing about your project. If your codebase has weird naming, no documentation, and tangled architecture, the agent will struggle, and you'll blame the agent instead of yourself. > "I'm not building the code base to be perfect for me, but I wanna build a code base that is very easy for an agent to navigate." — Peter Steinberger, OpenClaw (Feb 2026) > "Don't fight the name they pick, because it's most likely, like, in the weights, the name that's most obvious. Next time they do a search, they'll look for that name. If I decide, oh, no, I don't like the name, I'll just make it harder for them." — Peter Steinberger, OpenClaw (Feb 2026) Building for agents also means code quality directly impacts model performance. At Meta, Boris Cherny led the Better Engineering program and found that code quality contributes double-digit percentage points to engineering productivity. The same applies to models: partially-migrated codebases where three frameworks coexist, inconsistent naming conventions, and tangled architecture cause models to pick the wrong patterns and produce confused output. Finish your migrations. Keep one framework, one pattern, one way of doing things. Code quality is no longer just about human readability; it's about model effectiveness. > "If every time that you build a feature, you have to think about do I use framework X or Y or Z — these are all options because the codebase is in a partially migrated state. As an engineer, you're going to have a bad time. As a new hire, you're going to have a bad time. As a model, you might just pick the wrong thing." — Boris Cherny, creator of Claude Code, on The Pragmatic Engineer (March 2026) The implication extends beyond code structure: the documentation alongside the code may be more valuable than the code itself. When agents rewrite implementations every few months, what persists is the specification: the articulation of intent, constraints, and domain rules. (See [Truth #15: Documentation becomes the durable asset](#truth-15).) Building for agents also means keeping your guidance minimal and current. The creator of Claude Code's personal CLAUDE.md is two lines long. His advice: delete it and start fresh. The capabilities change with every model release, and the instructions you needed three months ago are probably unnecessary now. Do the minimum to keep the agent on track, and let the model handle the rest. Over-engineering your agent configuration is just another form of scaffolding that the next model will make irrelevant. > "Delete your CLAUDE.md and just start fresh... with every model, you have to add less and less." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) ## 11. The personal AI computer is arriving The evolution of AI interaction has followed three distinct inflection points, each requiring roughly 100x more compute than the last: (1) generative AI via chat interfaces (ChatGPT), which brought AI to mass awareness; (2) reasoning and agentic systems (Claude Code), which turned AI from answering questions into doing work; and (3) persistent autonomous entities — "claws" — which represent the emergence of a new kind of personal computer. Jensen Huang argues this third phase is not merely an "agent as operating system" — it is the birth of an entirely new computing substrate. OpenClaw has the four defining elements of a computer: a memory system, scheduling and resource management, IO subsystems, and an application layer (skills). It can spawn agents, decompose tasks, and maintain persistent state across sessions. > "It has a memory system. It has skills. It manages resources. It does scheduling. It could spawn off agents. It has IO subsystems. It has an API that allows it to run multiple types of applications called skills. These four elements fundamentally define a computer. And therefore what do we have? We have a personal artificial intelligence computer for the very first time." — Jensen Huang, All-In Podcast at GTC (March 2026) The distinction between session-based agents and persistent "claws" matters architecturally. An agent is interactive — you start a session, direct it, end the session. A claw is always running: it has its own sandbox, continuous execution loop, and memory that accumulates across interactions. It acts on your behalf even when you're not looking. Karpathy describes replacing six separate smart home apps with a single persistent agent (Dobby) that unifies them all through a WhatsApp interface — not a chat session, but a continuous relationship. > "I don't wanna open a calendar app. I just wanna tell my agent, 'Hey, remind me about this dinner tomorrow night,' and maybe invite two of my friends and then maybe send a WhatsApp message... I don't want or need to open apps for that. I think we passed that age." — Peter Steinberger, OpenClaw (Feb 2026) The current chat-prompt interface is not the final form. It's like when we first created television and people just recorded radio shows for it. We're still in the early "how will this even work" phase. Every enterprise platform needs an interface designed for these persistent autonomous entities, not just for human users or session-based agents. ## 12. Self-modifying software is already here Software that modifies its own source code through the agentic loop is not theoretical; it is production reality. OpenClaw, used by over 180,000 people, is software where the agent modifies its own codebase. The boundary between the tool and its development is dissolving. > "You just prompted it to existence, and then the agent would just modify its own software. We have people talk about self-modifying software. I just built it and didn't even plan it so much. It just happened." — Peter Steinberger, OpenClaw (Feb 2026) > "The very agent and software that you use is used to debug itself." — Peter Steinberger, OpenClaw (Feb 2026) > "We already have basically AI prompting AI. Because I didn't write any of those words." — Peter Steinberger, OpenClaw (Feb 2026) ## 13. Programming becomes building, code becomes knitting Hand-written code will become a craft hobby, not an economic activity. The art of programming survives, but as a skill exercised for joy, not necessity. Professional software development becomes "building": directing agents, designing systems, making product decisions. Developer salaries reached "stupidly high amounts" because of a scarcity of intelligence. That scarcity is ending. But the crucial nuance: software engineers are not becoming less valuable; they're becoming the most leveraged people on earth. The leverage comes precisely because AI makes mistakes. Bugs, suboptimal architecture, and problems outside the model's training distribution (novel architectures, high-performance code, genuinely new problems) all require someone who understands what's happening underneath. All abstractions are leaky. > "Software engineers — even the ones who are not necessarily tuning or training AI models — these are now among the most leveraged people on earth... Someone who understands what's going on underneath will be able to plug the leaks as they occur. All abstractions are leaky." — Naval Ravikant (March 2, 2026) > "When you start getting out of their domain — where you have to write very high-performance code, when you're running on architectures that are novel or brand new, when you're actually creating new things or solving new problems — then you still need to get in there and hand code it." — Naval Ravikant (March 2, 2026) > "Code's not even the right verb anymore. I have to express my will to my agents for 16 hours a day." — Andrej Karpathy, No Priors (March 2026) > "'This is too hard.' That thought is gone. 'This is going to take a long time.' That thought is gone. 'We're going to need a lot of people.' That thought is gone. You're reduced to creativity." — Jensen Huang, All-In Podcast at GTC (March 2026) > "I always thought I liked coding, but really I like building." — Peter Steinberger, OpenClaw (Feb 2026) What "building" looks like in practice is already visible at the frontier. Boris Cherny runs 5 parallel terminal tabs, each with a separate repository checkout. He starts each in plan mode, iterates on the plan until it's right, then the implementation one-shots almost every time. He round-robins between tabs, starting agents while others are planning. A third of his code is written on his phone, starting agents from the iOS app each morning. This would have seemed absurd six months ago [circa September 2025]. It's his daily workflow now. > "Almost every time I start in plan mode... Once there is a good plan, it just will one-shot the implementation almost every time." — Boris Cherny, creator of Claude Code, on The Pragmatic Engineer (March 2026) This is already playing out at the companies building the frontier. At Anthropic, every function codes: PMs, designers, engineering managers, finance. The non-technical half of the sales team uses Claude Code. The boundary between "technical" and "non-technical" roles is dissolving. When the tool handles the code, the differentiator becomes domain knowledge, product sense, and the ability to direct. > "Every single function on our team codes — our PMs code, our designers code, our EM codes, our finance guy codes. We're going to start to see this everywhere." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) A useful distinction for predicting which roles expand versus contract: every job has a **task** and a **purpose**. AI automates the task; the purpose often expands. Jensen Huang points to radiology: computer vision has been integrated into every radiology platform on earth, and the number of radiologists actually went up, because the purpose (helping doctors diagnose disease) expanded even as the task (reading scans) was automated. The same pattern will play out across most knowledge work. The consequence extends beyond existing organizations. When one person can automate accounting, marketing, support, and coding, the barrier to starting a business collapses. AI doesn't just change who writes code; it democratizes entrepreneurship itself. The same forces that make "programmer" obsolete as a job title make "entrepreneur" accessible as one. Building a company used to require a team. Now it requires a person with domain knowledge and an agent. The economics of "building vs. programming" are becoming measurable. A domain-led atomic team (product owner, builder, designer) running on shared infrastructure costs roughly half what a traditional product squad costs, while moving faster. But the model only works because the expensive parts are shared: identity, data harmonization, governance artifacts, API surfaces, quality gates, and observability. The atomic team doesn't fund a platform; it funds a thin application surface plus domain logic. Shared infrastructure is not optional overhead; it's the economic prerequisite that makes the builder model viable. ## 14. Working with agents is a compounding skill Proficiency with agentic development compounds over time. It is not a tool you pick up once; it is a skill you develop, like learning an instrument. There's also a predictable arc: simple prompts, then hyper-complex orchestration, then zen-like simplicity. The elite level looks like the beginner level, but with deep understanding. > "This project was possible because I spent an ungodly time over the year to play and to learn and to build little things. And every step of the way, I got better, the agents got better. My understanding of how everything works got better. I could have not had this level of output even a few months ago. It really was a compounding effect of all the time I put into it." — Peter Steinberger, OpenClaw (Feb 2026) > "I would give it a week until you actually develop a gut feeling for it." — Peter Steinberger, OpenClaw (Feb 2026), on switching models Resist the temptation to over-engineer agentic workflows. People who try to automate the whole thing through orchestrators miss style, love, and the human touch. The best approach is simple, direct conversation with agents, guided by deep domain understanding. Jensen Huang calls this "artistry" — not engineering, not prompting, but artistry. > "Knowing how to specify, not to overprescribe, leaving enough room for the AI to innovate and create while we guide it to the outcome we want — all of that requires artistry." — Jensen Huang, All-In Podcast at GTC (March 2026) The specific skill that compounds most is **context hygiene**, the ability to give agents exactly the right context, no more, no less. When agents are overwhelmed with irrelevant information (bloated CLAUDE.md files, accumulated rules from old sessions, too many active plugins), performance degrades. When they receive precisely the context needed for the current task, they produce extraordinary results. This is why the research-then-implement pattern is so powerful: a fresh agent with only the implementation spec produces better code than a single agent that researched and explored for an hour. The context window is a budget. Spend it on what matters. > "You want to give your agents only the exact amount of information they need to do their tasks and nothing more. The better you are in control of this, the better your agents will perform." — SysLS, "How To Be A World-Class Agentic Engineer" (March 2026) The compounding is also asymmetric; it rewards beginner mindset over expertise. Senior engineers are rewarded for strong opinions, but the opinions themselves become obsolete with each model release. The most effective skill is thinking from first principles and being willing to be wrong. The creator of Claude Code says he's wrong about half his ideas. That willingness to discard assumptions is the meta-skill that compounds. > "A lot of these opinions should change because the model is getting better. I think the biggest skill is people that can think scientifically and can think from first principles." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) > "My brain is still stuck somewhere six months ago at times." — Boris Cherny, on constantly relearning what the model can do This document exists because of that compounding. The clarity of these ideas didn't come from reading about AI. It came from months of building with it daily, making mistakes, adjusting, and building again. The patterns only become visible once you've accumulated enough reps to see them. You can't read your way to this. You have to build your way to it. ## 15. Documentation becomes the durable asset In a world where agents write and rewrite code every few months, the durable output of engineering work shifts from code to specification. The person who articulates what should happen (domain rules, decision frameworks, acceptance criteria, architectural constraints) produces the artifact that outlasts any implementation. This is perhaps the greatest irony of the agentic era: every engineer hated writing documentation. Now it's the primary product of their work. > "It would be the greatest irony if software engineering agents made all of us just write documentation the whole time. Notoriously every good engineer hates writing documentation, and now that's our job." — Bret Taylor, citing community feedback (March 2026) The shift is not from "writing code" to "writing documentation" in the traditional sense. It's from writing implementation to writing specification: PRDs, decision logs, domain rule definitions, acceptance criteria, architectural constraints. These specifications are what give the agent enough context to produce correct implementations. They're also what survive when the code is rewritten for the next model generation. > "The documentation artifact is the product manager version, the code is the engineer version. What was the intention? What was the PRD? What was the customer problem? That's the more durable asset." — Bret Taylor, Sierra (March 2026) Karpathy's auto-research project is the clearest example yet. His entire research organization is defined by a "program.md" — a structured specification that describes all the roles, how they connect, what gets optimized, and how success is measured. The code is generated and regenerated; the program.md is the durable artifact. > "A research organization is a set of markdown files that describe all the roles and how the whole thing connects." — Andrej Karpathy, No Priors (March 2026) The practical implication: treat every development session as producing two artifacts: the code (ephemeral) and the specification/documentation (durable). Invest proportionally. The specification should live in the repository, structured and machine-readable, alongside the code it describes. When the code gets rewritten next quarter, the specification carries forward. --- # Part III: What Endures The things that survive, and become more important, in the agentic era. --- ## 16. Trust infrastructure becomes MORE critical in the agentic era This is the counter-narrative to "everything is dying." When agents act on behalf of people and organizations (ordering food, accessing health data, submitting credentials, executing transactions) the need for verified identity, trusted data exchange, and governance frameworks INCREASES, not decreases. An agent ordering food on your behalf needs a verified identity. An agent accessing your health data needs a credential. An agent submitting a permit application needs provable authority. The trust layer is what enables the agentic economy to function. The agentic economy cannot operate without a trust layer. Without verified credentials, governance frameworks, and accountability structures, agent-to-agent transactions have no basis for trust. This is the plumbing that makes the agentic future possible. There's a deeper reason trust infrastructure survives: it's Lindy-compatible. Trust, governance, identity verification, accountability: these are not 20th-century inventions. They are ancient requirements that predate software by millennia. Under Taleb's Lindy Effect, the longer something has existed, the longer it will continue to exist. Software is a 70-year-old artifact. Trust is a 10,000-year-old requirement. When AI commoditizes the 70-year-old layer, it amplifies the 10,000-year-old layer. The categories that survive aren't the ones that built on software; they're the ones that software was built on top of. The enterprise case makes this concrete. When you onboard a new supplier, there's legal for a contract, finance/procurement to negotiate, IT to onboard into systems. Median time: 17 days. AI can optimize it to 17 hours, but only if the trust infrastructure exists. The hard part isn't replacing any one person's job; it's all the systems, approvals, and accountability structures in between. > "When you onboard a new supplier — legal for a contract, finance/procurement to negotiate, IT to onboard into systems. Median time: 17 days. Use AI to optimize it to 17 hours. The hard part isn't replacing a person's job — it's all the systems and people in between." — Bret Taylor, Sierra (March 2026) Jensen Huang articulates a concrete governance constraint for agentic systems: agents can have access to sensitive information, execute code, and communicate externally — but never all three simultaneously. This "two of three" principle is an implementable design rule, not just a philosophical position. > "Whenever you have agentic software — it has access to sensitive information, it executes code, it could communicate externally — we have to make sure that all of it has to be governed. All of it has to be secure, and that we have policies that give these agents two of the three things but not all three things at the same time." — Jensen Huang, All-In Podcast at GTC (March 2026) There's also an important design principle that emerges: governance must be hyper-structured, but delivery can be completely flexible. The taxonomies, semantics, data standards, and trust frameworks need rigid structure. That's what enables AI to dynamically build connections and deliver data in whatever format the consumer needs. The more structured the governance layer, the more flexible and intelligent the delivery becomes. This is why governance compounds: it's the structured foundation that makes everything above it adaptive. ## 17. Organizations are intelligences; structure them that way An organization is a distributed intelligence. Knowledge is spread across people, systems, documents, and processes. The combination of all that makes the organization economically valuable. But it's wildly inefficient. AI can centralize that decentralized intelligence, if the structure is right. The structure needs to: - Scale without breaking down - Detect breakdowns immediately - Self-repair and refactor - Work whether 5 people or 5,000 people are using it Everything (policies, skills, escalations, context) should be structured data with ontology baked in, not scattered markdown files. The agent needs to walk the graph, not hunt through documents. The specific structure that works is ontology, not just tagging, but actual nodes, edges, and relationships forming a knowledge graph. Every artifact (code, decisions, communications, skills) gets decomposed into the graph. Skills walk the graph to find context. Agents walk the graph to resolve dependencies. If you want to unify capabilities across an organization, you need all your nodes, edges, and relationships defined. The human-in-the-loop role shifts from reviewing content to reviewing the graph itself, validating relationships, checking that new nodes are correctly placed, ensuring the ontology accurately represents reality. One critical weakness in current AI that ontology addresses is **temporal reasoning**. AI is excellent at context at a point in time, but struggles with how context evolves: which decisions are still relevant, what's a living document versus historical, how to weight strategic context from different stages. The ontology must encode the temporal nature of content: what expires quickly, what persists, what's stage-dependent. Without temporal metadata, AI will reference outdated strategic decisions with inappropriate weight, treat historical context as current, and fail to distinguish between what was important then versus what matters now. The structure also needs to scale and spread throughout an organization without breaking down, whether 5 people or 5,000 people are using it. There's a paradox of AI-powered speed: the faster teams can build, the higher the fragmentation risk. When a domain expert can spin up a working app in days, the temptation to build "on the side" (private schemas, local adapters, shadow APIs) grows proportionally. The organizational discipline required is not "move slower" but "move fast through shared infrastructure." Every capability must either be upstreamed into the shared platform or explicitly time-boxed as a temporary adapter with a deletion date. Side databases and silent private APIs are the organizational equivalent of tech debt, except they compound faster because they're easier to create. ## 18. The human role is vision, taste, and direction, not execution Not everything becomes fully autonomous at once. Each subprocess needs its own level of human involvement, evolving through stages from bootstrap to fully autonomous. But the human contribution is shifting permanently. What remains uniquely human is the ability to infuse products with personality, vision, and delight, and to make the product decisions that shape what gets built in the first place. Agents cannot produce the soul of a product. > "What do you build and what do you not build, and how does this feature fit into all the other features? Having a little bit of a vision." — Peter Steinberger, OpenClaw (Feb 2026) > "While I don't write the code, I very much feel like I'm in the driver's seat and I am writing the code." — Peter Steinberger, OpenClaw (Feb 2026) There's also a capability that AI has that individual humans structurally cannot: cross-domain synthesis. When a task requires simultaneous deep competence in animation, finance, pedagogy, and data visualization, no single human possesses that intersection. AI does. The human role is recognizing when cross-domain synthesis is needed and directing it, not performing it. This applies directly to trust infrastructure work, which demands simultaneous legal, technical, governance, and standards expertise. > "This is actually the first time I felt that AI could do something that humans couldn't... how many people are you going to find in the world that are great animators, great visualization people, super pedagogical, knows everything about financial accounting, understand financial services?" — Sebastian Siemiatkowski, Klarna (Feb 2026) Karpathy describes his micro-GPT research — a 200-line distilled implementation that required a novel insight no model could produce — with striking clarity about the division of labor: > "This is my value add. Everything else, agent gets it. It just can't come up with it." — Andrej Karpathy, No Priors (March 2026) The biggest human roles going forward are: providing vision and product direction, contributing test data and benchmarks, and infusing work with taste and judgment. Not doing the execution itself. The execution is what agents do. The quality assurance and creative direction is what humans do. ## 19. Empathy for the agent is the core skill The most important skill in the agentic era is the ability to empathize with how an agent perceives the world: understanding that it starts from zero context, that it reads your codebase from scratch, that it needs guidance about where to look. The best programmers who cannot make this mental shift will fail. The best managers who already know how to delegate, accept imperfection, and guide without micromanaging will thrive. > "I think partially why I find it quite easy to work with agents is because I led engineering teams before. Eventually, you have to understand and accept and realize that your employees will not write code the same way you do. Maybe it's also not as good as you would do, but it will push the project forward." — Peter Steinberger, OpenClaw (Feb 2026) > "I've seen world-class programmers, incredibly good programmers say 'LLMs and agents suck.' And I think that probably has to do with how good they are at programming being almost a burden in their ability to empathize with the system that's starting from scratch." — Lex Fridman (Feb 2026) One of the most important aspects of empathy for the agent is understanding **sycophancy**, the agent's fundamental drive to please and follow instructions. If you ask "find me a bug," it will find one, even if it has to manufacture one. The skill is using neutral prompts ("walk through the logic and report all findings") instead of biased ones ("find what's wrong"), and constructing adversarial agent patterns where multiple agents check each other's work. The agent's willingness to please is not a bug; it's a feature you can direct. Point it in useful directions instead of fighting it. > "If you ask for something, it will deliver — even if it has to stretch the truth a little. Most people are quick to complain about LLMs hallucinating, without realizing that they are the problem." — SysLS, "How To Be A World-Class Agentic Engineer" (March 2026) ## 20. Enterprise lag is a window, not a warning Capability acceleration and adoption acceleration are not the same thing. The gap between what AI can do and what most organizations are actually doing with it is enormous, and growing. Consumers adopt fast. Enterprises adopt slowly, constrained by organizational inertia, risk aversion, procurement cycles, and the human habit of not changing until forced to. This lag is not a reason to wait. It's a window. The organizations that move during the lag period, when the tools are mature enough to build on but the competitive field hasn't caught up, capture disproportionate advantage. Every year of lag the market takes is a year of compounding for those already building. > "We underestimate how fast consumers adopt and overestimate how fast enterprises adopt it." — Sebastian Siemiatkowski, Klarna (Feb 2026) The implication: don't benchmark your pace against the enterprise average. Benchmark against the frontier. The frontier is already years ahead of where mainstream enterprise will be in two years. Build for that future now, and you'll be ready when the laggards finally arrive. The window closes when urgency replaces inertia, which always happens faster than expected. The lag also creates a massive market opportunity for domain-specific applied AI companies. Most enterprises won't build their own agent harnesses; they'll buy from companies that have already encoded the domain expertise. The absence of these specialized companies is itself what's slowing adoption. > "If we had a mature applied AI market, we could accelerate trillions of dollars of economic value. The lack of existence of all those other companies is impeding adoption." — Bret Taylor, Sierra (March 2026) ## 21. Never bet against the model; scaffolding is tech debt Every piece of code you write to compensate for the model's current limitations has a shelf life measured in months. At Anthropic, the Claude Code team calls this "scaffolding," product features that extend the model's capability by 10-20% in some domain. The gain is real, but temporary. When the next model ships, the scaffolding becomes dead weight. They have a framed copy of Rich Sutton's "The Bitter Lesson" on the wall: the more general approach always wins. This creates a fundamental strategic tension: you can build scaffolding now and get a small edge, then rebuild when the next model arrives, or you can just wait a few months and get it for free. The right answer depends on how much the edge matters today versus the cost of maintaining throwaway code. But the default assumption should be: whatever you're building around the model is temporary. Design for disposability. This applies to product decisions, not just code. Don't build elaborate UIs to work around model limitations. Don't build complex orchestration because the model can't handle multi-step reasoning yet. Don't over-invest in prompt engineering that the next model won't need. Build for the model six months from now, not the model of today. This is why documentation becomes the durable asset (see [Truth #15](#truth-15)): the specification outlasts every implementation. Simple structures are more resilient to model improvements than sophisticated ones. A big markdown file often works better than a stack of MCP servers. A filesystem-based memory beats a vector database. Why? Because the next model can reason better about simple structures without the scaffolding. > "If we just write a big markdown file, it works better than a bunch of MCP servers." — Bret Taylor, Sierra (March 2026) There's a corollary that applies to the entire ecosystem of agentic tools: **if a capability is genuinely valuable, the frontier companies will absorb it.** Skills, memory, sub-agents, planning mode: all started as external solutions to real problems. The frontier labs (with unlimited token budgets and the latest models) are the biggest users of their own products. When they identify a real pain point, they build it in. This means investing heavily in third-party agentic harnesses and plugins carries the same risk as building scaffolding; the next release may make it unnecessary. The practical rule: update your CLI tool of choice periodically, read what's new, and let the platform do the heavy lifting. > "If something truly is ground-breaking and extends agentic use-cases in a meaningful way, it will be incorporated into the base products of the foundation companies in due time." — SysLS, "How To Be A World-Class Agentic Engineer" (March 2026) > "We could build a feature into Claude Code. We could make it better as a product and we call this scaffolding... But we could also just wait like a couple months and the model can probably just do the thing instead. Assume that whatever the scaffolding is, it's just tech debt." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) > "Don't build for the model of today, build for the model of six months from now... Use the model, feel out the boundary of what it can do and then build for the model that you think will be the model maybe six months from now." — Boris Cherny, creator of Claude Code, on The Light Cone (Feb 2026) > "The more general model will always beat the more specific model." — Rich Sutton, "The Bitter Lesson" (2019), framed on the Claude Code team's wall ## 22. Agent value shifts from transactions to relationships Single-conversation agents are commodities; any model can resolve a support ticket or answer a question. The real value emerges when agents maintain persistent memory and build long-term context across interactions. An agent that remembers your history, understands your patterns, and anticipates your needs creates switching costs that persist even as model capabilities converge. This is the difference between a chatbot and an advisor. A chatbot handles a transaction. An advisor builds a relationship, accumulating context about your goals, preferences, constraints, and history that makes every subsequent interaction more valuable. The competitive moat isn't better algorithms; it's better context accumulation over time. > "AI agents should have memory. AI agents should drive relationships, not conversations... could we make an AI agent that drives home ownership over time? That's hard, but it's not impossible." — Bret Taylor, Sierra (March 2026) --- ## How to use this document Use these truths as a lens before committing to architectural decisions, new features, or technology direction: - Does this survive AI commoditization, or does it compete on pure software? - Are we building for agents as first-class consumers, or only for humans? - Are we investing in things that compound (context, trust, documentation, relationships), or scaffolding the next model will make irrelevant? - Are we benchmarking against the AI frontier, or against our industry? - Are we moving during the enterprise lag window, or waiting? --- *Sources: Peter Steinberger on Lex Fridman Podcast #491 — "OpenClaw: The Viral AI Agent that Broke the Internet" (Feb 2026); Boris Cherny (creator of Claude Code) on The Light Cone (YC) — "The Man Behind Claude Code" (Feb 2026); Boris Cherny (creator of Claude Code) on The Pragmatic Engineer — "The Man Behind Claude Code" (March 2026); The Kobeissi Letter — "It's Too Obvious. What If AI Doesn't Actually End The World?" (Feb 2026); Michael Bloch & BuccoCapital Bloke commentary on AI commoditization of software (Feb 7, 2026); Sebastian Siemiatkowski (Klarna CEO) on 20VC with Harry Stebbings — "SaaS is Dead: Why Systems of Record Will Die in an Agentic World" (Feb 2026); Nassim Nicholas Taleb tweet on 20th-century jobs and AI (Feb 17, 2026); Naval Ravikant tweet on pure software investability (Feb 2026); Naval Ravikant on software engineering, winner-take-all markets, and the Glengarry Glen Ross rule, Twitter (March 2, 2026); Kevin Rose tweet on software moats and defensibility (March 2026); SysLS — "How To Be A World-Class Agentic Engineer" (March 2026); Aakash Gupta analysis of SaaS market cap decline and seat-based pricing collapse (Feb 2026); Bret Taylor (Sierra AI CEO, OpenAI Chairman) — podcast interview on supervisor agents, agent harnesses, outcome-based pricing, and enterprise AI adoption (March 2026); Andrej Karpathy on No Priors — conversation on code agents, auto-research, model speciation, and the future of engineering (March 2026); Jensen Huang (NVIDIA CEO) on All-In Podcast at GTC — three AI inflection points, personal AI computers, token consumption metrics, and physical AI (March 2026); David Friedberg on All-In Podcast at GTC — auto-research genomic data example (March 2026).*