The Brightest Minds Converge on the End of Hardcoded Software
March 14, 2026
Ten voices. One week. One architecture. A defense CEO, the Pentagon CTO, a Sequoia partner, and a Rhodes Scholar mathematician all described the same missing layer in AI.
Ten voices. One week. One architecture.
The CEO of a $360 billion defense company, the Pentagon's most senior technology official, and a 24-year-old Rhodes Scholar dropout all described the same missing layer in AI last week. None of them were talking to each other. Neither were the Sequoia partner, the former Salesforce co-CEO, or the Box CEO who said essentially the same thing on the Latent Space podcast the same week.
Ten separate people, in the span of a few days, arrived at the same conclusion from entirely different starting points. They used different vocabulary. They were addressing different audiences. But the structural observation underneath was identical, and I have not seen anything like this convergence in a decade of building infrastructure for regulated financial markets.
Karp told the truth on camera
Alex Karp does not do small talk. So when the Palantir CEO said this week, on camera, that "the LLMs on the battlefield are only useful if powered inside an Ontology, and that's our Ontology," he was making a precise claim about how military AI actually works.¹
Not how it should work, but how it works right now, today, inside the operational infrastructure that the United States uses to project force.
Palantir built a $360 billion company on a specific insight: intelligence systems need a structured representation of the world they operate in before they can do anything useful. They call this the ontology. Every entity, every relationship, every constraint, mapped and enforced. The language model sits on top. The ontology sits underneath. Without the ontology, the model is generating plausible text. With it, the model is operating inside a verified representation of reality.
The constraint Palantir faces is that building each ontology requires Forward Deployed Engineers who spend months embedded with the client. Every new deployment is a custom job. That works for governments. It does not work for the other 200,000 enterprises that need the same infrastructure.
Musk announced the autonomous enterprise
The same week, Musk unveiled Macrohard, a joint Tesla-xAI project designed to emulate the operations of entire software companies using AI. The system pairs Grok as the reasoning layer with a Tesla-built agent as the fast executor, watching screen activity and taking actions autonomously.²
Musk framed it as "emulating the function of entire companies." Set aside the characteristic overstatement. The directional claim is serious: AI agents will not assist individual workers. They will operate entire workflows end to end.
But companies do not operate in isolation. They contract with each other. They are bound by regulations that cross jurisdictional boundaries. They transact under documented terms that both parties must follow. When Company A's autonomous agent needs to execute a transaction with Company B's autonomous agent, both agents need to operate under shared rules that neither can override.
Nobody in the Macrohard announcement addressed who enforces those rules. Because the agents do not know the rules exist. The rules live in contracts, compliance manuals, and operational procedures that no one has turned into executable infrastructure.
The Pentagon named the risk
Days later, Emil Michael, the Pentagon's Chief Technology Officer and Under Secretary of Defense for Research and Engineering, raised an alarm about Claude, Anthropic's language model, being embedded across defense contractor systems through Palantir's platform. Michael told CNBC that Claude's built-in "constitution" carries a "different policy preference" that could "pollute" the military supply chain.³
The technical detail that matters here is not the sentience debate. It is the supply chain argument. The Pentagon identified that a probabilistic system embedded in deterministic operational infrastructure creates structural risk. The AI has its own "constitution," as Anthropic describes it. But the defense establishment has operational procedures, safety protocols, and rules of engagement documented across thousands of pages. The question is whose rules the AI follows when those two frameworks conflict.
This is not a hypothetical, and it is worth sitting with the fact that the Department of Defense described, in public, the exact failure mode that occurs when AI operates without a constraint layer extracted from the organization's own documented rules.
Taylor redefined the unit of productivity
Bret Taylor, former co-CEO of Salesforce and chairman of the OpenAI board, said this week that "the atomic unit of productivity in AI is a process, not a person." He described how a 17-day workflow spanning legal, finance, procurement, and IT could collapse into hours when an AI agent executes across all departments simultaneously.⁴
Taylor's framing matters because it shifts the conversation from replacing workers to compressing workflows. But it also surfaces a problem he did not address. When an agent executes simultaneously across legal, finance, procurement, and IT, it must follow the documented rules of each department. Legal has contract terms. Finance has accounting standards. Procurement has vendor agreements. IT has security policies. Each department's rules are documented somewhere. None of those documents have been turned into executable constraints that an agent can operate within.
Compress a 17-day process into 17 hours without enforcing documented rules across every silo, and you have produced 17 hours of unchecked compliance exposure at machine speed.
Sequoia called the market
Julien Bek, Partner at Sequoia Capital in London, published "Services: The New Software" on March 5. His core thesis: for every dollar spent on software, six dollars are spent on services. The next trillion-dollar company will be a software company masquerading as a services firm. If you sell the tool, you are racing against the models. If you sell the work, every model improvement makes your service faster, cheaper, and harder to compete with.⁵
Bek distinguishes between copilots (tools for professionals) and autopilots (systems that deliver the outcome directly). He argues that the autopilot model starts where outsourcing already exists: the company has accepted external execution, there is a defined budget line, and the buyer is already purchasing outcomes.
This matches what we learned running a full-service treasury operation in regulated European debt capital markets. We did not sell compliance software to financial institutions. We ran the operation. Documentation, validation, settlement. The technology was the engine. The service was the product. Bek's thesis describes what we built before he wrote the article.
But the autopilot model has a constraint that Bek's framework does not address. When you sell the outcome in a regulated environment, you are legally liable for the accuracy of that outcome. The autopilot needs a rule book it cannot override, extracted from the actual documentation that governs the operation. Without that, you are selling outcomes you cannot guarantee.
Levie quantified the governance gap
Aaron Levie, CEO of Box, speaking on the Latent Space podcast ("Every Agent Needs a Box," March 5, 2026), made the scale problem concrete: "Whether you think the number is 10x or 100x, we're going to have some order of magnitude more agents than people."⁶
Then he described what happens: "There's going to be just incredibly spectacularly crazy security incidents that will happen with agents, because you'll prompt-inject an agent and find your way through the CRM system and pull out data you shouldn't have access to."
His conclusion: "No matter what, there's going to need to be a layer that manages the data they have access to and the workflows they're involved in. This is the new infrastructure opportunity in the era of agents."
Levie is describing a governance layer. But governance requires knowing what the rules are. You cannot enforce access controls you have not defined. You cannot manage workflows whose constraints have not been formalized. And the rules that govern enterprise workflows are not sitting in a database. They are sitting in contracts, policy documents, regulatory filings, and operational procedures that nobody has extracted into executable form.
Sivulka saw the factory floor
George Sivulka, CEO of Hebbia and a16z portfolio founder, published the most complete articulation of the problem this week. His framing: "We've swapped the motor; we have not yet redesigned the factory."⁷
He traces the analogy to the 1890s electrification of textile mills. Factories replaced steam engines with electric motors but kept the same floor layout, the same power distribution architecture, the same workflow design. For thirty years, the new motors produced almost no increase in output. It was not until factories were completely redesigned around the capabilities of electricity that the productivity gains materialized.
Sivulka's essay outlines seven pillars of what he calls "Institutional Intelligence." Three of them connect directly to the infrastructure question.
On determinism: "Institutional-grade intelligence must be defined, deterministic, and auditable in the work it does." He explicitly contrasts individual AI (nondeterministic, prompt-driven) with institutional AI (deterministic, process-driven, with "predictable checkpoints, steps, and processes").
On enforcement: "The most important agents inside organizations will not be 'yes-men' but disciplined 'no-men' that interrogate reasoning, surface risks, and enforce standards." He lists the consequential applications: AI compliance, AI auditors, AI board members. All constraint functions.
On process: He names Palantir as "one of the first true process engineering companies" and argues that process engineering, not software engineering, will become "arguably the most important technology in the near term."
Hong proved the architecture in mathematics
Carina Hong is a 24-year-old Rhodes Scholar and Stanford JD/PhD dropout who founded Axiom seven months ago to build an AI mathematician. Her system, Axiom Prover, solved all twelve problems on the 2025 Putnam exam, making it the first AI to clear the full set. The proofs are not approximate or probabilistic. They are formally verified by machine, with no human intervention required.⁸
What makes Axiom relevant to the infrastructure question is not the mathematics but the architecture underneath it.
Axiom works because it bridges two layers: an informal reasoning layer (the language model, which generates candidate solutions in natural language) and a formal verification layer (Lean, a programming language for mathematical proofs that can deterministically verify whether a proof is logically correct). The language model proposes, the formal system verifies, and only verified outputs pass through.
When asked whether this architecture generalizes beyond mathematics, Hong was direct: "Do you need a Lean equivalent for each one of those domains as you expand?" Her answer: "I think so."
That is the critical constraint. Axiom achieves 100% correctness because mathematicians spent decades building Lean as a complete formal verification system for mathematical logic. The formal language existed before the AI arrived. Without it, the language model would be generating plausible but unverifiable solutions — which is exactly the hallucination problem everyone else is trying to solve.
In mathematics, that layer is Lean. In software, it could be a strongly typed language like Rust. In business operations, it does not exist. The rules that govern enterprise workflows, encoded in contracts, compliance documentation, operational procedures, and regulatory filings, have never been extracted into a formal system that AI can verify against.
Axiom did not solve the hallucination problem in general. It solved it in one domain where the formal verification infrastructure already existed. Extending that architecture to any other domain requires someone to build the equivalent formal layer first. For business operations, that means extracting the documented rules of an organization into a deterministic, executable ontology.
What the convergence tells us
Ten people arrived at the same conclusion in the same week, working independently, speaking to different audiences, using different vocabulary. A defense CEO, a tech billionaire, a Pentagon official, a former Salesforce co-CEO, a Sequoia partner, a Box CEO on an a16z podcast, a Hebbia CEO reposted by Marc Andreessen, a Rhodes Scholar mathematician who built formally verified AI in her domain, and the current and former CEOs of the two largest enterprise workflow companies.
Strip away the different vocabularies and what remains is a single architectural claim: probabilistic AI cannot operate in high-stakes environments without a deterministic constraint system underneath it. Karp calls that system an ontology and builds it manually with engineers. Taylor assumes it will emerge from process compression but does not say who builds it. Musk ignores it entirely. Levie names it as the governance layer but has no formalized rule base to govern against. Sivulka calls it the redesigned factory but does not specify the operating system. Hong proved that the architecture produces 100% correct output in mathematics, then identified that for business operations, the formal verification layer has never been built.
None of these ten people used the same word for what they were describing. But having spent a decade building exactly this type of infrastructure, my reading of the convergence is specific: the constraint system they are all circling is an ontology — a structured and executable representation of every entity, relationship, rule, and validation that governs how an organization operates. It must be extracted from the documents that already define those operations, because the rules already exist. They are written down in contracts, policy manuals, compliance frameworks, and regulatory filings, and they have never been turned into infrastructure that AI can operate within.
The technology to extract those rules automatically, encode them as deterministic constraints, store operational data against them, execute validations in real time, and expose the results to agents and humans has already been built and tested in production.
In regulated European debt capital markets, this architecture has processed hundreds of millions in transaction volume with 99.98% accuracy across more than 100,000 validations and zero compliance failures over three years of continuous operation.
Why this changes everything
There is a reason this convergence matters beyond the immediate question of whether AI can be made reliable. It matters because the ontology does not just sit alongside existing enterprise software. It threatens to replace it.
Consider what companies like ServiceNow and SAP actually sell. Bill McDermott, CEO of ServiceNow and former CEO of SAP, went on CNBC this same week to argue that AI gives great advice but cannot execute, and that the "last mile" between AI reasoning and operational action belongs to workflow software.⁹ His successor at SAP, Christian Klein, makes the same argument: "In the enterprise world, where we are setting up our agents, you need 100% accuracy. We are running these business processes, we know the rules and workflows, we have the data."¹⁰
Both are saying that hardcoded workflow platforms remain indispensable in the age of AI.
They are wrong about why, and right about what is at stake.
The value inside ServiceNow and SAP is not the orchestration logic, the user interface, or the integration layer. The value is the business rules that consultants spent years hardcoding into those systems. When a ServiceNow workflow routes a procurement approval through the correct chain of sign-offs, the intelligence is not in the routing engine. It is in the knowledge of who must sign, under what conditions, at what thresholds, governed by which policy. That knowledge was extracted manually from operational documents by human consultants and then frozen into software configuration.
If you can extract business rules automatically from the source documentation, validate them, keep them current as the documentation evolves, and execute operations against them in real time, then the container those rules used to live in becomes a commodity. The orchestration layer does not disappear overnight, but its strategic value migrates — from the software to the rules it encodes, and from the vendor who configured them to the infrastructure that extracts and maintains them at machine speed.
The ten voices in this article are not just describing a missing layer for AI. They are describing the beginning of a structural shift in how enterprises encode and execute their own operational logic. For sixty years, that logic has lived inside software platforms, maintained by consultants, and understood only by the people who configured it. The moment it becomes extractable, portable, and machine-verifiable, the relationship between enterprises and their software changes in a way that does not reverse.
Within 24 months, at least one Fortune 500 company will have its operational rules extracted from source documentation, encoded into a machine-executable ontology by dedicated infrastructure, and its compliance workflows running against that ontology rather than against the platform that originally housed them. When that happens, it will surface a question that most enterprises have never had to ask: does anything in your technology stack actually know how your operation works, in a form that a machine can verify?
The companies that can answer yes will deploy AI agents with structural trust. The rest will continue struggling with the adoption barriers that Citadel Securities mapped three weeks ago¹¹ — the same barriers that ten of the brightest minds in technology pointed at independently last week.
Robert Koller is the Founder and CEO of SynapseLayer, building automated ontology infrastructure for enterprise AI. Learn more at synapselayer.ai.
Footnotes
- Alex Karp, CEO of Palantir. AIPCon 9 remarks, March 12, 2026. [source]
- Elon Musk, Macrohard / Digital Optimus. Reuters coverage, March 11, 2026. [source]
- Emil Michael, Pentagon CTO. CNBC Squawk Box, March 12, 2026. [source]
- Bret Taylor, co-founder Sierra, Chair OpenAI. Cheeky Pint Podcast, March 2026. [source]
- Julien Bek, Partner Sequoia Capital. "Services: The New Software," March 5, 2026. [source]
- Aaron Levie, CEO Box. Latent Space podcast, "Every Agent Needs a Box," March 5, 2026. [source]
- George Sivulka, CEO Hebbia. "Institutional AI vs Individual AI," March 12, 2026. [source]
- Carina Hong, CEO Axiom. MAD Podcast with Matt Turck (FirstMark), March 2026. [source]
- Bill McDermott, CEO ServiceNow (former CEO SAP). CNBC Squawk on the Street, March 13, 2026. [source]
- Christian Klein, CEO SAP. TIME interview, Aug 2025; Stratechery interview on enterprise AI, May 2025. [source] [SynapseLayer analysis]
- Citadel Securities. Frank Flight, "The 2026 Global Intelligence Crisis," Feb 24, 2026. [source] [SynapseLayer analysis]