The Centaur

Chapter 1: The Pilot Model — Human + AI × Agents

"Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process."
— Garry Kasparov, Chess Grandmaster

The Operating Model: Humans Direct, AI Amplifies, Agents Execute

The pilot model is simple: one human (or team) directs AI-powered agents toward a clear mission. The human provides vision, judgment, and approval. The AI translates intent and orchestrates execution. Autonomous agents work in parallel, each bounded by mission documents, reporting results through the Glass Box.

This isn't theoretical. It was first proven in freestyle chess tournaments (2005-2008), where human-machine teams — sometimes called "centaur teams" after Kasparov's term — consistently outperformed both the best humans AND the best machines alone. As Harvard Data Science Review notes, these players identified relative advantages between themselves and their chess programs, even when those programs were individually superhuman — a result that held through the freestyle era of 2005–2008, though, as we will see, its lesson would outlive the fact itself.

The insight is profound: it's not about human OR machine. It's about the process that combines them. A weak human with a machine and a good process beats a strong human with a machine and a poor process. The process is everything.

Before these ideas had a name, we were building something very close to this model at a national telecommunications carrier. We orchestrated machine learning for outage prediction using weather bureau data, linear programming for real-time truck-roll optimisation of six thousand field technicians, augmented reality visors and tablets that used vision learning to detect hardware types, loose cables, LED states, and physical signs of faults — all coordinated through a single operational lens. We didn't call it a cockpit. But that's what it was: human operators directing AI systems that commanded field agents, with transparent reporting flowing back through the Glass Box. The productivity gains were substantial. But more importantly, it proved the principle — the process of combining human judgment with machine capability and distributed execution is what produces extraordinary output. The technology was less mature than what's available today. The pattern was identical.

The Research Evidence

The 5–10x Multiplier — and Where It's Heading

This model reflects what Peter Diamandis calls the "demonetisation and democratisation" of complex capabilities — the pattern where exponential technologies make previously exclusive tools universally accessible. Salim Ismail's Exponential Organisations framework — digitise, dematerialise, democratise, demonetise — describes the organisational architecture required to harness rather than resist this collapse.

The pilot model extends beyond a human-AI pair. When a human pilot works alongside an AI co-pilot commanding fleets of autonomous agents, the multiplication effect is extraordinary.

One pilot, an AI, and a fleet of bounded agents can reach an output a room of people could not. The honest evidence forms a ladder. Measured in controlled trials: 25–55% task-level gains (GitHub Copilot RCT; MIT Sloan found 27–39% for junior developers) — yet traditional enterprises capture only ~10% at the organisational level, because complexity eats the difference. That gap is this series' entire argument. Measured at AI-native organisations: 5–20x revenue per employee versus the $150–250k traditional software benchmark — Midjourney at $5M+, Cursor at ~$3.3M, Lovable at ~$2.2M per person — and build speeds to match: according to Anthropic's own engineers, Claude Cowork was built in about ten days, with effectively all of its code written by Claude Code (employee accounts, not an audited figure). Demonstrated in cases: order-of-magnitude results on properly decomposed work, from MAKER's million error-free steps to Knuth's "Claude's Cycles." The fair counter-evidence: METR's randomised trial found experienced developers were actually slower with early-2025 AI on mature, entangled codebases — proof that the multiplier is not automatic. It is conditional on exactly the workflow discipline this essay describes.

This isn't science fiction. Industry leaders are already seeing it:

Who	Prediction / Evidence
Sam Altman (OpenAI)	"10-person billion-dollar companies" coming soon
Dario Amodei (Anthropic)	One-employee billion-dollar startup possible by 2026
Y Combinator (W25)	25% of startups had codebases 95% AI-generated
Midjourney	~40 employees, $300M revenue, $5M+ revenue per employee
Base44	Founded solo; a handful of employees by its ~$80M acquisition (Wix, 2025)

Why Agents Don't Add Communication Overhead

Brooks' Law states that adding people to a late project makes it later, because communication overhead scales quadratically:

Team Size	Lines of Communication
5 people	10
15 people	105
100 people	4,950

The Standish Group's CHAOS Reports — three decades of project data covering tens of thousands of IT initiatives — confirm what this mathematics predicts: small projects succeed at roughly ten times the rate of large ones. Projects that stay small in scope, team size, and duration consistently deliver; projects that scale beyond a handful of people and a few months consistently fail on at least one metric of time, cost, or quality. Jeff Bezos codified the same insight as Amazon's "two-pizza team" rule; J. Richard Hackman's fifty years of team research converged on four to six people as the optimal unit. The mathematics of coordination explains why.

In the AI era, the penalty for ignoring this mathematics has become catastrophic. When a single contributor's output was measured in thousands of dollars, the coordination tax of a sixth or seventh team member was a manageable overhead. When AI amplifies individual output by five to ten times, that same coordination tax is measured in millions of lost productivity. As analyst Nate B Jones argues, volume is no longer the scarce resource — correctness is. A twenty-person team generates more AI-assisted volume, but shared context degrades. A five-person team optimises for correctness, because every piece of AI output passes through human brains with enough shared context to catch meaningful errors.

The insight at the heart of the ORBIT methodology — the subject of Essay V: Agents don't add communication overhead. They work in isolated contexts, bounded by mission documents, reporting results to one Pilot through the AI. Twenty agents = zero additional communication connections.

From "Doing" to "Directing"

Microsoft's vision for the "Frontier Firm": "Where we once said, 'I send emails,' 'I write documents,' 'I create pivot tables,' we'll soon say, 'I create and manage agents.'"

McKinsey's view: "As agents take on execution, people will increasingly define goals, make trade-offs, and steer outcomes."

This is exactly what the ORBIT methodology delivers: the Pilot defines vision. The AI translates intent. Agents execute in parallel. The human directs; the AI army does.

The Pilot Model Across Functions

The pilot model works identically regardless of domain — only the mission and the agents change:

Domain	The Pilot	The AI	The Agents
Software Dev	Lead developer	Code synthesis, architecture review	Coding agents, test agents, review agents
Design	Creative director	Design generation, brand compliance	Layout agents, asset agents, variant agents
Marketing	Marketing lead	Content strategy, audience analysis	Writing agents, analytics agents, campaign agents
Finance	CFO / Controller	Modelling, anomaly detection	Reporting agents, compliance agents, forecast agents
Sales	Sales lead	Research, personalisation	Outreach agents, proposal agents, analysis agents
Enterprise	CEO / COO	Cross-functional synthesis	Department-specific agents, monitoring agents

---

Chapter 2: The Unified Mission Cockpit

"There's an important difference between hiding information and making it inaccessible."
— Unix Philosophy

One Cockpit, Not Fifty Dashboards

Most "simplification" tools face a fatal tradeoff: they hide complexity but also hide accountability. When something goes wrong, you can't see what happened. When auditors ask questions, you have no answers. When experts need to dive deep, they can't.

The Mission Cockpit takes a fundamentally different approach:

Glass Box, Not Black Box

This is the defining principle of the entire architecture. Every AI system claims capability. The dividing line is transparency.

Black Box AI (the industry default)	Glass Box AI (the Centaur model)
Inputs and operations not visible	All parameters known, conclusions traceable
Can't explain which features led to outputs	Every decision traced to mission documents
Auditors can't verify compliance	Complete audit trail: what, why, when, who approved
Experts locked out of details	Progressive disclosure: any depth available on demand
EU AI Act compliance: difficult retrofit	Governance-ready by design

Progressive Disclosure: Simplicity at Every Level

Progressive disclosure — initially showing only the most important options, then revealing specialised options on request — is one of the best-attested patterns in interface design: usability research consistently finds it improves learnability, efficiency, and error rates (Nielsen Norman Group).

The cockpit implements this through layers: the AI provides the simple conversational interface. A status layer shows peripheral awareness. Quality scores surface issues. But underneath, every agent action is logged, every decision is traceable, every AI choice can be inspected.

Natural Language as the Universal Interface

Nielsen Norman Group identified natural language interaction as "the first new UI interaction paradigm in 60 years" — and the productivity evidence in Chapter 1, from NN/g's task studies to Brynjolfsson's new-hire findings, rests on exactly this interface.

The AI-first approach achieves what traditional tools cannot: a common language that unites all stakeholders.

• Interns can ask questions in plain English and get answers they understand
• Developers can describe what they want without translating to tool-specific syntax
• Product managers can discuss features without learning technical jargon
• Executives can understand progress without decoding status reports
• Clients can engage directly with the product vision

The complexity is minimal and robust by design. The specialised work that requires deep expertise is covered by AI agents in the background — abstracted away — so that the interface returns to the simple, common language that unites all teams.

Why This Matters for Governance

As AI regulation tightens globally, the ability to explain AI decisions isn't optional — it's legally required.

Regulatory Framework	Key Requirements
EU AI Act (Aug 2026)	High-risk AI must be "sufficiently transparent to enable users to interpret the system's output"
U.S. State Laws (2025)	1,100+ state AI bills introduced; major states passing disclosure requirements
SOC 2 / GDPR	Data processing transparency, right to explanation
Article 14 EU AI Act	Effective human oversight with measures matching risks

The Glass Box is governance-ready by design: all decisions trace to mission documents. Human approval gates for protected documents. Every AI action logged with context. "AI proposes, human approves" as architectural principle.

Why the Human Stays in Command

There is a deeper reason the human stays in command — and it is not that the pair computes better. Be honest about the chess story that opened this essay. In 2005, human-plus-machine "centaur" teams genuinely beat the strongest engines alone. That is no longer true: as the engines grew superhuman, the human hand became a liability in pure, fully-specified games, and the machine alone now wins (Krakowski, Luger and Raisch, Strategic Management Journal, 2023). And chess is not unusual — from Paul Meehl's Clinical versus Statistical Prediction (1954) to a 136-study meta-analysis by Grove and colleagues (2000), simple models beat or match expert human judgement roughly 94% of the time. A 2024 meta-analysis in Nature Human Behaviour (Vaccaro, Almaatouq and Malone; 106 experiments) sharpened the point: on average, human-AI combinations performed worse than the best of human or machine alone — genuine synergy appeared mainly in creation tasks, rarely in decisions. If we justified human command by claiming humans decide better, we would forfeit that command the moment the machine improved. And it always improves.

So we ground it elsewhere, on foundations that do not erode. Decisions carry values, and values require an owner. Someone must be answerable for a choice — to customers, to colleagues, to regulators, to the public. A machine can compute a recommendation; it cannot be accountable for one. This is the principle behind the "meaningful human control" doctrine (Santoni de Sio and van den Hoven, 2018) and behind human-oversight law such as Article 14 of the EU AI Act. It is why "AI proposes, human approves" is not a workflow convenience but a moral architecture — the same principle that, at enterprise scale, will govern who may see and own your data.

---

Chapter 3: The View System — Multiple Lenses on One Reality

"Reality is singular, but perspectives are many."

What Is "Reality" in an Organisation?

Before examining the views, we must answer a fundamental question: what is "reality"?

Everything else is a lens — a projection of these elements combined. Even "external" things collapse into data: an API is its documentation plus its behaviour (captured as logs). User behaviour is analytics data. Regulations are legal documents. If something matters to the system, it manifests as data. If it doesn't manifest as data, we can't see it through any lens anyway.

What Does a Lens Do?

A lens is a projection that shows reality from a specific angle:

• A Requirements Lens doesn't just show code — it shows code in relation to stated requirements
• A Compliance Lens doesn't just show data — it shows data in relation to regulatory obligations
• A Financial Lens doesn't just show transactions — it shows transactions in relation to budgets and forecasts

Every lens asks: "Show me this slice of reality."

The Core Views

Different moments call for different perspectives:

View	What It Shows	When You Need It
Horizon View	Where we came from → Where we are → Where we're heading. Mission/Vision, key milestones, the "why this matters" narrative	Strategic planning, team alignment, board presentations
Lab View	Active experiments. Hypotheses being tested with explicit success/failure criteria	Innovation cycles, R&D, market testing
Flow View	Current work in progress. Blockers, dependencies, who's working on what	Daily operations, sprint execution
Evidence View	Metrics that matter (not vanity metrics). Product-market fit indicators. User feedback synthesis	Performance reviews, investment decisions
Journey View	Achievements unlocked. Milestones reached. Team growth. Celebrates successes AND valuable failures	Motivation, retrospectives, team culture
Pivot View	Major direction changes documented. Why we pivoted, what evidence drove it. Evolution of understanding over time	Strategic reviews, institutional memory

Enterprise Lenses: Beyond the Six Views

For the enterprise, the view system extends dramatically. The same principle — different lenses on one reality — applies across all functions:

Functional Lenses — each department gets a view filtered through their domain:

Lens	Shows	Synthesises From
CFO Lens	Everything in financial terms. Revenue, cash flow, unit economics, budget variance	ERP, billing, CRM, payroll, forecasting
CTO Lens	System health, tech debt, architecture fitness, incident patterns	Monitoring, code repos, CI/CD, security tools
HR Lens	Headcount, engagement, skill gaps, attrition risk	HRIS, engagement surveys, performance data, Slack sentiment
CMO Lens	Pipeline, brand health, content performance, channel ROI	Analytics, CRM, social platforms, ad systems
CEO Lens	Strategic alignment across all functions. "Are we doing what we said we'd do?"	ALL of the above, synthesised

Cross-Cutting Lenses — where the real enterprise value emerges:

Lens	Shows	Why It's Powerful
Customer Lens	One customer's full journey across every system	Sales + support + product + billing + NPS, unified
Compliance Lens	Regulatory exposure across the enterprise	Legal + finance + ops + data, cross-referenced
Competitive Lens	Everything known about competitors	Sales battle cards + win/loss + product telemetry + research

Temporal Lenses:

Lens	Shows
Retrospective	"What happened in Q4 across the whole business? What were the real causes?"
Real-time	"What is happening right now that needs attention?"
Predictive	"Based on all patterns, what is likely to happen next quarter?"

Lenses Compose

This is where it becomes transformative. Lenses combine to answer questions that currently require cross-functional war rooms:

Each of these questions currently takes days or weeks to answer — requiring data from multiple systems, meetings with multiple teams, and manual synthesis. With the View System on a unified Knowledge Fabric, they become queries.

---

Chapter 4: Living Documents — Boundaries That Evolve Through Evidence

"The most dangerous words in business are 'we've already committed to this direction.'"

The Boundary Paradox

A profound paradox sits at the heart of innovation: you need boundaries to make progress (can't explore infinite directions), but rigid boundaries prevent discovery (might miss the real opportunity).

Most methodologies resolve this paradox poorly:

Approach	The Problem
Waterfall	Locks boundaries too early. Teams execute plans that should have been abandoned.
Agile	Often lacks boundaries entirely. Teams iterate endlessly without strategic coherence.
Traditional Strategy	Documents treated as sacred. Updated annually at best. Discoveries that challenge the vision are dismissed.

History is littered with examples: Kodak's strategic vision didn't include cannibalising film. Blockbuster's architecture didn't accommodate streaming. Nokia's product definition couldn't embrace touchscreens. The boundaries that enabled their success became the constraints that ensured their failure.

The Resolution: Safe Experimentation

The ORBIT methodology resolves the boundary paradox through a principle: experiments operate in isolation, where they are free to explore without risk — free to modify not just implementation, but the foundational assumptions themselves.

The key insight: Mission documents are the starting point for exploration, not the boundary that constrains it. Within the isolated experiment, AI agents can question anything — including the product vision and strategic direction. But nothing changes in reality until the human pilot approves.

What This Enables

Hypothesis-testing at the vision level: An experiment could explore: "What if our target market is actually enterprise, not consumer?" The AI modifies the mission documents to reflect this alternative vision, builds a prototype consistent with it, and presents the complete package for human evaluation.

Mission-bound or goal-attaining: Experiments can be bounded to mission documents (refine within constraints) or exploratory (question everything). Constraints can be set: time, duration, cost, quality, or target outcome.

Parallel exploration: Multiple experiments can simultaneously explore different strategic directions — "What if we went upmarket?" "What if we went freemium?" "What if we pivoted to a platform model?" Each returns with a complete package. The human pilot compares parallel directions and chooses which to commit.

The documents tell the discovery story: Over time, the version history of mission documents becomes a rich narrative. Each evolution is annotated with the experiment that proposed it, the evidence that supported it, and the human decision that approved it.

Safe-to-Fail Probes at Every Level

Level	Traditional Approach	ORBIT Approach
Mission	Sacred, unchangeable	Can be questioned in isolated exploration
Vision	Updated annually at best	Continuously tested via parallel experiments
Architecture	Locked after initial design	Evolved based on experimental evidence
Strategy	Quarterly planning cycles	Hypotheses tested daily
Features	Sprint-based delivery	Continuous improvement with quality gates

The revolutionary implication: Nothing is sacred except the human's right to decide. Everything — from code to architecture to vision to mission — can be explored, questioned, and experimentally revised. But nothing changes in reality until the human pilot approves. Maximum exploration with maximum safety.

---

Chapter 5: The Architecture — CRUD + AI and the Blessed Stack

"If you have a database, an AI, and APIs to the outside world — do you need Mailchimp? HubSpot? Salesforce? Answer: No. You need none of them."

The First-Principles Architecture

Strip any enterprise function to its essence and you need exactly four components:

The Quality-at-Source Principle

There is a deeper principle at work: the quality of outputs is fundamentally determined by the quality of inputs.

Mission documents are not merely documentation. They are the agreements that bind the work — the authoritative WHAT and HOW that every line of code generated by the AI, every decision made by an Agent, and every quality check is built against. But they bind the work, not reality: the contract is the current best hypothesis, written down — firm enough to build on, humble enough to revise. It is spec-anchored, not spec-as-scripture; when the evidence and the contract disagree, it is the contract that updates.

Barry Boehm's research demonstrated that fixing a defect after deployment costs up to 100x more than fixing it at the requirements stage (Boehm's classic estimate; the exact multiplier is contested, the direction is not). Current software projects waste 40-50% of effort on avoidable rework, and 80% of that stems from just 20% of defects (industry rules-of-thumb) — almost all originating in unclear requirements.

This echoes Sakichi Toyoda's revolutionary principle — jidoka, quality at the source — which became a pillar of the Toyota Production System. If the contracts are right at the source, downstream execution has a far greater chance of being right.

The Blessed Stack: Radical Simplification Through Opinionated Choice

Beyond the AI multiplier, there's another transformational advantage: the Blessed Stack — an opinionated, pre-integrated technology foundation that eliminates the endless complexity of "build vs. buy" decisions. It is a default, never a cage: every layer can be swapped, and the value is a sensible starting point, not a locked one.

This approach, pioneered by Spotify ("Golden Path") and Netflix ("Paved Road"), delivers quantifiable benefits:

The Blessed Stack prevents tech debt accumulation by design. Companies that allow unconstrained technology choices inevitably accumulate incompatible systems requiring expensive transformation programmes.

Pattern Recognition: The Discovery Principle

The Mission Cockpit isn't just an execution tool — it's a discovery engine. When AI has visibility across all your data and systems, it doesn't just do things faster — it surfaces things you'd never find.

Discovery patterns:

Pattern	Example
Cross-domain correlation	"Healthcare customers have 3x higher retention but you've never targeted them deliberately"
Anomaly detection	"This metric changed trajectory 3 weeks ago — before the symptom became visible in reports"
Opportunity surfacing	"Based on usage patterns, these 15 customers are ideal expansion candidates"
Risk identification	"This combination of factors preceded churn in 8 of your last 10 lost accounts"

Pattern recognition works because the AI has access to ALL the data — sales, product usage, support, renewal, financial — and can synthesise across it. In a world of siloed tools, these patterns are invisible. In a unified Glass Box, they're obvious.

---

Essay III Summary

THE MECHANISM IS CLEAR. What does it unlock?

ESSAY IV: THE COLLAPSE