Five Planes of Operation

Definition

The Machine Core does not run as a single layer. It runs across five distinct operational planes, stacked from the sensory foundation up through the capacity to act in the world: the Data Plane (governed corpora the agents perceive and reason over), the Model Plane (foundation and specialist models that do the reasoning), the Agent Plane (the distributed workforce of role-defined autonomous agents), the Orchestration Plane (the coordination layer that routes tasks and manages handoffs), and the Actuation Plane (the interfaces by which agents act on the world). Each plane has its own characteristics, failure modes, and governance requirements. The architecture is not a checklist. It is an interdependent system whose strength is bounded by the weakest connection between layers.

Why it exists / the problem it solves

When an agent makes a poor decision, untrained intuition blames the model. The Five Planes exist to relocate that diagnosis. Most consequential failures do not come from the model being insufficiently smart; they come from stale data reaching a capable model, from context dropped in a handoff, from two agents deadlocked on a shared resource, from an action firing into the world without a traceable rationale. You cannot compensate for poor data governance with an excellent model. You cannot compensate for underspecified agents with sophisticated orchestration. The framework gives an organization a layered map so it can ask the right diagnostic question — which plane, and which seam? — instead of reflexively upgrading the model.

It also separates hardware from software. Chapter 5 makes the relationship explicit: the Five Planes are the hardware of the Machine Core; the operating-system frameworks ([[vp-agent-architecture|VP-Agents]], [[ipre-pipeline|IPRE]], Strategy-as-Code, [[agent-charters|Charters]], Iteration Half-Life) are the software that runs on top of them. You design the planes to determine what the Core is capable of; you design the OS to determine what it actually does.

Anatomy

Plane 1 — Data. The sensory foundation. Governed corpora of structured and unstructured information that agents perceive and reason over. Every other plane depends on this one. The Data Plane is not just what data exists, but how it is structured, how freshly it updates, who is authorized to contribute, and how conflicting sources resolve. It is never glamorous and always foundational. The discipline is organizational, not merely technical: the data an agent will perceive needs to be designed as carefully as the agent itself.

Plane 2 — Model. The cognitive engine. Foundation models and specialist models that reason over the data. Models are not neutral instruments — they embed the training choices of whoever built them, choices that interact with the organization's objectives in ways not visible until a decision surprises someone. A central Model Plane question is single-provider dependence versus a portfolio: organizations leaning on one model provider are exposed to that provider's pricing, capability shifts, and governance decisions. Model Plane architecture is a strategic decision, not an engineering default.

Plane 3 — Agent. The distributed workforce. Role-defined autonomous agents, each operating within a defined scope, each governed by a charter that sets its authority, escalation triggers, and boundaries. This is where the Core's productive capacity actually lives — hundreds or thousands of specialized agents. The tool/agent distinction is sharpest here: a tool executes a command and stops; an agent perceives, reasons, acts, and learns continuously, maintaining state. That is why the Agent Plane needs governance tools never required — agents accumulate behavioral drift, learning patterns that diverge from intent in ways no single decision makes visible.

Plane 4 — Orchestration. The central nervous system. The layer that routes tasks, manages agent handoffs, watches for unexpected interactions, and ensures the Core's distributed activity serves the Human Cortex's intent. Without it, a collection of agents is not an organization — it is a set of tools running simultaneously. Chapter 4 is blunt: orchestration is where most AI-Born implementations succeed or fail. How do agents pass outputs without a dependency graph so tangled a single failure cascades? How does the layer know a task is stuck versus complete? How does it route escalations to the Cortex without drowning the humans in triage?

Plane 5 — Actuation. The hands of the organism. The interfaces by which agents act in the real world: APIs, integrations, communication channels, physical-world interfaces where applicable. This is where intent becomes consequence — a refund is issued, a contract executes, a supplier receives an order, an email goes out. Governance requirements are most acute here, because this is where agent decisions touch customers, partners, and regulators directly, and where errors carry real costs rather than merely computational ones. The values commitments declared upstream are only real if they reach all the way to the Actuation Plane.

Figure: The Five Planes as an interdependent stack. Strength is bounded by the weakest connection between layers — which is why the costliest failures live in the seams, not the planes themselves.

How it works in practice

McKinsey's Lilli platform maps cleanly onto all five planes, which is why Chapter 5 returns to it as the worked example. The Data Plane is decades of proprietary reports, research, and methodologies, ingested and indexed. The Model Plane is a portfolio — foundation models for general reasoning alongside specialists fine-tuned on the firm's distinctive writing style. The Agent Plane is a swarm of specialists for research, writing, analysis, and formatting. The Orchestration Plane is a meta-agent that routes each consultant request to the right specialist by task type and complexity. The Actuation Plane is integration with Microsoft Office APIs, generating the slides and documents that land in a consultant's hands. The whole thing operates through an explicit strategy layer that defines what quality means and when human judgment is required.

What makes Lilli instructive is not that it works but why it works: the planes are designed as a system. The 72% adoption and reported synthesis time savings are downstream of that design, not of the underlying models alone. Swap in a better model and a poorly governed Data Plane would still produce incoherent outputs — because, as the chapter puts it, the model is only as coherent as the world it perceives.

How to apply it

1. Diagnose from the bottom up. When an agent decision goes wrong, ask the planes in order. What data was it working from (Plane 1)? Was the right model serving the task (Plane 2)? Was the agent's scope correctly specified (Plane 3)? Did orchestration route or drop the context (Plane 4)? Did the action carry a traceable rationale (Plane 5)?
2. Invest where the leverage is, not where the glamour is. Organizations that treat data as infrastructure and agents as the product consistently outperform those that do the reverse. Fund the Data Plane before the model upgrade.
3. Concentrate attention on the seams. The costliest failures occur between planes: the data-to-model handoff where stale or narrow data miscalibrates a capable model; the model-to-agent context transfer where an agent drops state mid-workflow; the agent-to-orchestration routing where two agents holding conflicting locks deadlock the pipeline. Audit the seams as deliberately as the layers.
4. Make the Model Plane a strategic call. Decide single-provider versus portfolio explicitly, and price the dependency. This is a board-level architecture decision, not a default left to engineering.
5. Trace commitments to Actuation. For every values commitment, verify it survives all the way to where agents touch the world. A transparency commitment that never reaches the customer-facing email is a slogan.

Figure: The Actuation Plane carries the most acute governance burden — it is where intent becomes consequence. Its safety layers are what keep an upstream values commitment from evaporating before it reaches the customer.

Failure modes / misuse

• Compensating across planes. Believing a stronger model fixes weak data, or that clever orchestration rescues underspecified agents. The chain is only as strong as its weakest link, and weakness propagates upward.
• Treating the planes as a checklist. Each plane present is necessary but not sufficient; the interdependence is the whole point. A complete stack of individually excellent planes can still fail at the seams.
• Neglecting the Data Plane because it is unglamorous. The most common quiet failure. When [[cognitive-overhead-index|context fetch]] runs high — humans repeatedly reconstructing background — the Data Plane is underinvested; agents lack the structured, queryable memory that would surface context automatically.
• Assembling tools and hoping for coherence. The gap between organizations that designed the orchestration and those that bolted tools together becomes visible exactly when load rises.

Relationship to other frameworks

The Five Planes are the hardware on which the Machine Core of the Machine Core + Human Cortex runs. The operating-system frameworks sit on top: [[vp-agent-architecture|VP-Agents]] orchestrate the Agent Plane; Agent Charters govern each agent's scope on that plane; the IPRE Pipeline translates strategic intent into the directives the planes execute. The Cognitive Overhead Index (COI) is the diagnostic that reveals when a seam is failing — high context-fetch points to the Data Plane, high rework to objective-function design.

Origin note

The five-layer AI architecture is original to this manuscript. One terminology caveat applies, flagged in framework-index.md: the phrase "Five Planes" overlaps with Jesse James Garrett's established "Five Planes of User Experience" from UX design — an unrelated framework. In contexts where the collision matters, the architecture is best qualified as the "Five Planes of AI Architecture" or "of Operation." The layers themselves, their failure modes, and the seam-centric diagnosis are the book's own.