The New Scarcity

On February 4, 2026, roughly one year after he made the term famous, Andrej Karpathy retired vibe coding. The man who coined it killed it himself. His replacement framing: agentic engineering. The defining quote: “The new default is that you are not writing the code directly 99% of the time. You are orchestrating agents who do and acting as oversight.” Weeks later he was clearer still: coding agents “basically didn’t work before December and basically work since.” The biggest change to his coding workflow in two decades.

Markets had been pricing this in for weeks. On January 12, 2026, Anthropic launched Claude Cowork as a research preview, a non-coding agent product that made AI execution of knowledge work concrete in a single demo. The market reaction unfolded over the following weeks, accelerating sharply when Anthropic published 11 enterprise plugins for Cowork on January 30. The 48-hour wipeout that followed approached $285 billion in SaaS market cap. Jefferies trader Jeffrey Favuzza coined the term SaaSpocalypse to describe the moment. Cumulative losses across the sector reached $1 trillion or more by April, with some estimates approaching $2 trillion. Atlassian reported its first-ever decline in enterprise seats.

When markets reprice an entire sector and the cultural anchor of the prior era retires the term he made famous within days of each other, you are not in an evolutionary moment. You are in a phase transition. The substrate has changed. Vocabulary, companies, investors, and ecosystems are catching up at different speeds, investors and ecosystems slowest.

Call it the divide. The word matches the experience. Pre-divide: the world that produced great companies between roughly 2010 and 2024, anchored by per-seat pricing, scaled by hiring, defended by switching costs. Post-divide: different physics, different team shapes, different defensibility. The substrate shifted operationally with Claude 3.5 Sonnet in mid-2024, accelerated when OpenAI released o1-preview in September 2024, became measurable through 2025 (Claude Code GA, AI-native burn multiples diverging), and reached public consensus in early 2026 with the SaaSpocalypse and the agentic engineering reframe.

What changed

Three data points sketch the divide.

Workflow. Anthropic shipped Cowork in approximately a week and a half, largely written by Claude Code itself. Boris Cherny, who created Claude Code, stated in late January 2026 that 100% of his code (not just edits, all of it) has been written by Claude for two-plus months, shipping 22 PRs in a single day on this basis. Anthropic’s company-wide figure is 70 to 90%. Dario Amodei at Davos in January 2026: “I have engineers within Anthropic who say I don’t write any code anymore. I just let the model write the code, I edit it.” The trajectory is six to twelve months out from end-to-end automation, on the timeline the company building the substrate is already living.

Markets. The pricing model collapse is sectoral, but not monolithic. Per-seat pricing share dropped from 21% to 15% in a single year. 47% of SaaS companies are exploring outcome-based models. 85% have adopted or are exploring usage-based pricing, up from a much smaller base two years ago. Intercom’s Fin AI charges $0.99 per resolved ticket and grew from $1M to $100M+ ARR on that single pricing move. ServiceNow Now Assist crossed $600M ACV in late 2025. The split is between systems-of-record (Salesforce, Microsoft, ServiceNow: absorbing agent-pricing, structurally protected by data moats) and bolt-on point solutions (exposed, structurally cannibalizable). The mechanism beneath the collapse is the consumption-surface collapse: agents do not need dashboards, navbars, or seats.

Per-seat pricing was a tax on human attention. The tax base is shrinking.

Unit economics. AI-native companies are operating at burn multiples materially below traditional SaaS peers, with capital efficiency that is reshaping how investors evaluate the category. Lovable hit $100M ARR with approximately 45 employees. Multiples bifurcated: legacy SaaS once trading at 18x revenue at the 2021 peak now sits in the 5-7x range.

The playbook is not dying gracefully. The substrate has changed.

Execution is a commodity now

Pre-divide, execution was the scarce resource. Engineering hours, design hours, copywriting hours, sales hours: the bottleneck. Capital was raised to buy execution. “Hire 10 engineers to ship faster” was a coherent strategy. Companies that could afford more execution capacity outshipped companies that could not.

That equation is broken.

Anthropic ships products in a week and a half. Lovable hits $100M ARR with approximately 45 employees. It operates at over $2 million revenue per employee, a benchmark virtually unheard of among early-stage companies a decade ago. When execution is no longer scarce, companies that hire to scale execution are buying a commodity at premium prices. Companies optimizing for the new scarcities outcompete them.

This is uncomfortable for the parts of the ecosystem whose value proposition was matching execution capacity to demand. Many founders, investors, executive recruiters, and large-team consultancies have not yet internalized that this has changed.

If execution is the commodity, what is the scarce resource now?

What is actually scarce

Six resources, each genuinely scarce in 2026, each impossible to download:

Creative reasoning. The model can answer almost anything you ask it. The scarce skill is asking the right thing. Knowing what is worth building, what frame to apply, what question to investigate. The bottleneck has moved from “can we execute on this idea?” to “is this the right idea, framed the right way, asked of the right substrate?”

Disciplined workflow. How to structure work so AI augmentation compounds rather than producing churn. Specific patterns: spec-driven development, reflexivity discipline, Pareto frontier discipline, adversarial model convergence. Production tooling shipped through 2025 (GitHub Spec-Kit, Tessl, OpenSpec, BMAD). Thoughtworks Tech Radar Volume 33 placed spec-driven development in Assess in November 2025 while flagging that poorly executed spec-driven workflows risk reverting to waterfall, the exact tension that defines the discipline gap. Tools commoditize fast; the disciplines stay scarce because they require deliberate cultivation.

Architectural thinking. Translating fuzzy problems into specifications agents can execute. The new core engineering skill is structure-first thinking: designing the boundaries, contracts, and verification strategy before implementation. AI does the implementation; humans own the architecture, quality, and correctness. When AI generates the implementation, the integrity of the architecture becomes load-bearing in a way it never was when humans wrote the code.

Decision quality. Models accelerate option generation; humans still bear decision-making weight. The cost of a bad decision compounds faster when execution is fast. A pre-divide team could afford a bad decision because executing on it took weeks; a post-divide team cannot, because they ship the bad decision the same afternoon. This raises the premium on judgment, not lowers it.

Taste. The ability to recognize good output, push back on bad output, iterate toward better. Generalists with taste outproduce specialists without it. Taste compounds with use; it cannot be hired in. Post-divide founders are increasingly evaluated on taste (by customers, collaborators, capital partners) because taste is the layer the AI cannot supply.

Wide reasoning with stacked depth. Pure generalism produces dilettantes; pure specialism in a fast-shifting environment produces obsolescence. The post-divide winning pattern is broad horizontal capacity to direct AI across domains, combined with stacked vertical depth where judgment must be trustworthy: architecture, customer reasoning, market sense, strategic decisions. The pre-divide org chart specialized humans into narrow lanes; the post-divide org chart consolidates lanes into people with breadth plus deliberate depth where decisions compound.

None of these are tool problems. All are discipline problems.

Where we came in

September 2024 is an inflection most people who built through it can locate precisely. OpenAI’s o1-preview and the maturing Claude 3.5 Sonnet substrate opened genuine research and design work to a cohort that had previously been locked out of building software at meaningful complexity.

The cohort is broader than “engineers who started using AI tools.” Architects of systems, product builders, domain experts in fields adjacent to software, designers exploring implementation directly, researchers building instruments for their own questions, generalists suddenly able to operate across all the specialties they had spent careers translating between.

What the cohort discovered, building in the new substrate, was that the substrate rewarded workflows, not tools. The tool layer was already commoditizing. Two practitioners with the same toolset produced wildly different output depending on the workflow methodology they brought: architectural thinking that anticipated agent execution rather than human execution, iteration loops measured in hours, not weeks.

Through 2025, the cohort built, without waiting for permission, funding, or category validation. The substrate had opened a door and they walked through it.

What needed to be built

Of the six scarcities, the one that compounds fastest into structural risk is architectural thinking. When AI generates the implementation, the architecture is where human judgment lives. Everything downstream (whether the code works, whether it’s safe to ship, whether it does what was intended) depends on whether the architecture can carry the weight. Pre-divide, the architecture could be wrong and the implementation could quietly correct it through human authorship. Post-divide, that buffer is gone. The architecture is load-bearing in a way it never was when humans wrote the code.

The cohort surfaced this first because it broke first. Specs become the source of truth, but who verifies that the implementation matches the spec? When the chain from intent to deployed code is reconstructed through agent execution rather than human authorship, the verification step that used to live in human review has nowhere to anchor. Engineering teams, enterprises, capital partners, and compliance officers ask different versions of the same question: “how do you know the code does what the spec said?” The question lives upstream of any of them.

This is the integrity problem the post-divide world creates. Enterprises will not deploy AI-generated code into production at scale without integrity infrastructure. The infrastructure does not yet exist as a category. It has to be built.

A reasonable counter: this looks like a feature, not a company. GitHub Advanced Security, Snyk, Datadog, and the major cloud providers are all building AI-code verification. The counter has weight but a structural constraint: any vendor whose distribution is their ecosystem cannot be the standards layer for code that crosses ecosystems. The integrity problem is upstream of where the code lives, who hosts it, and which scanner runs against it. It has to be built by someone whose distribution is the standard itself.

Idora is the integrity layer for AI-agent-led software delivery. Foundation Capital, Gartner, and the W3C have converged on a name for the category: context graphs, governed memory layers that capture decision traces and provenance for agent reasoning. Idora’s integrity graph is a context graph specialized for verification. Every node is a spec-derived determination; every edge is a traceable proof chain. General-purpose context graphs ground reasoning in organizational memory. The integrity graph grounds reasoning in specifications. Verification has a different bar than recall.

Substrate is what distinguishes infrastructure that endures from infrastructure that compresses. A model can replicate a thin technical service as a feature; it cannot replicate accumulated state, regulatory provenance, or architectural commitment that compounds with use. The integrity layer is the second kind.

The verification engine grounds in spec-first reasoning: every determination traces back to the specification it was derived from. The integrity graph compounds per push: each customer interaction densifies the substrate that future verifications draw against. Built for the world where agents write the code and humans need to trust the output.

How we built it

Idora is post-divide infrastructure built using post-divide methods. The how is part of the thesis.

Engineering at Idora applies the disciplines named above to the integrity-layer problem: spec-driven workflow, adversarial convergence on architectural decisions, reflexivity discipline against our own production, Pareto-frontier discipline against bloat. Architecture documentation, investor materials, strategic memos, and technical specifications are all produced through the same disciplined-workflow loop.

Distribution is agent-runtime-first. The Agent Skills standard (Anthropic-originated, December 2025; adopted by Microsoft, OpenAI, Cursor, Vercel, Supabase since) is becoming for procedural knowledge what npm became for code: the package-manager primitive for the agent layer. Idora ships to Anthropic’s Skills Directory and Vercel’s skills.sh from day one. The integrity layer needs to be where agents already look for capability, not where humans go to download a SaaS app. Agent-runtime distribution is not a future channel; it is the channel.

Monetization anchors on decisions delivered, not access granted. Every Idora call returns a determination an agent or human can act on; pricing reflects what was produced. The post-divide AI category is converging on outcome-based pricing (Intercom per resolution, Sierra per outcome, Salesforce Agentforce per conversation). Per-decision pricing is the integrity-layer specialization: the unit being priced is the determination itself. Per-seat licensing prices the human surface area of a system; this system does not have one.

Defensibility is compounding intelligence: the integrity graph densifies per push, switching cost grows monotonically with use, replication requires going back to push one. These are the substrate properties that distinguish payment infrastructure, data warehouses, and observability platforms as durable horizontal infrastructure. The integrity layer occupies the same reference class.

Idora runs on Idora. The integrity layer that grounds customers’ AI-generated code is the same layer that grounds ours: every push compounds our own graph, every release runs through our own decision surface, every drift between spec and deployment surfaces first in our own integrity graph. Integrity all the way down: structural, not rhetorical. It is how the system stays honest about its own claims.

What this means for capital and ecosystem

Many investors are operating with playbooks built for the prior era. They evaluate AI-native companies on the metrics that ranked SaaS companies, frame defensibility in switching-cost terms when the substrate is compounding intelligence, and apply per-seat pricing logic to companies whose value is delivered as decisions, not access. None of this is malicious. It is operationally stale. AI-native companies that accept this guidance produce the bifurcation the data already shows.

The strongest case against the divide thesis is not that it’s wrong but that it’s early. Many of the data points (Cherny’s 100%, the Cowork demo) sit at the leading edge of a distribution most of the industry will reach over the next 18 to 36 months. The structural argument holds either way.

Geographic ecosystems outside the Bay Area (Houston, Austin, Toronto) realize their potential when they adopt new creative engineering thinkers who work within augmented AI workflows at hyper efficiency. Joe Schmidt’s contribution to a16z’s Big Ideas 2026 names forward-deployed AI outside Silicon Valley as a defining 2026 opportunity, predicting that “the vast majority of the AI opportunity lives outside of Silicon Valley.” That thesis rewards ecosystems that adopt the new founder archetype. It does not reward ecosystems that recreate the old one with cheaper rent.

Execution is a commodity now. The new scarcity is creative reasoning, disciplined workflow, architectural thinking, decision quality, taste, and the wide-with-stacked-depth capacity to direct it all.

Idora exists because the integrity problem the divide creates has to be solved, and the cohort that lived through September 2024 surfaced it as a category that needed infrastructure. If you are a founder building in the new substrate, a capital partner who recognizes the divide, or an ecosystem builder positioning for what is coming, the conversation starts from the same recognition. Everyone who lived through it knows what changed. The question is what gets built now that it has.