> axis 02 · structure · what must I supply for that expectation to hold?

The AI-Human Engineering Stack

Five layers of cognitive engineering, two meta-functions that operate across them, and a user interaction model that ties them together.

Working effectively with AI requires more than writing good prompts. Each layer answers a different categorical question. Each depends on all layers beneath it. The gap between people who use AI effectively and people who struggle is rarely about prompt skill. It is almost always about the invisible layers, the ones the Stack is designed to make visible.

§1 system_diagram

Seven components, one system

Five layers stack bottom to top. Evaluation runs alongside every layer; the harness is the platform underneath. Each layer is harder to see, harder to specify, and harder to evaluate than the one below. The amber current climbs the dependency spine; the green pulses are evaluation reading each layer. Mark any layer weak and watch what happens above it.

Prompt

What to do

scope Per-sessionfails as Incoherent capability: individually plausible, collectively inconsistent

state: ok

Context

What to know while doing

scope Per-systemfails as Informed incompetence: good instructions, wrong or missing information

state: ok

Intent

What to want while doing

scope Per-organizationfails as Misaligned optimization: task done correctly, actual goal undermined

state: ok

Judgment

What to doubt while doing

scope Continuousfails as Confident wrongness: high metric performance, progressive real-world drift

state: ok

Coherence

What to become while doing

scope Per-ecosystemfails as Emergent misalignment: each agent correct, collective behavior wrong

state: ok

Harness

Where and how to do

scope Infrastructurefails as Capability mismatch: the platform constrains what the cognitive layers could do

state: ok

the dashed rail on the right is Evaluation: how to know while doing

scope Cross-layerfails as Silent failure: the system can't see what's breaking or where

state: ok

binding constraint: none marked. Mark a layer weak to find it.

§2 dependency_rule

The dependency rule

Each layer depends on all layers beneath it. A system cannot have strong intent engineering without adequate context engineering: agents can't make aligned tradeoffs without the information to see what they're trading off. A system cannot have strong judgment engineering without clear intent: agents can't know when to doubt themselves if they don't know what they're supposed to want.

Identify the lowest weak layer. That is the binding constraint. Improving any layer above it before fixing the layer below produces diminishing returns. Most architectures in 2026 have their binding constraint at context or intent engineering.

§3 interaction_model

The user interaction model

The stack is not a one-shot configuration, but a continuous interaction between the user and the layers, mediated by evaluation.

The user begins at the bottom, with prompt engineering. Every AI interaction starts with an instruction. But before the first prompt is written, the user has typically already made harness decisions (which tool to use) and may have configured context, intent, judgment and coherence. The stack is configured bottom-up but often set up in parallel.

The diagnostic loop

When output is unsatisfactory, the natural instinct is to rewrite the prompt. The Stack gives a more precise diagnostic.

Evaluate. Evaluation detects a problem, either autonomously (tests fail) or through human review (the output doesn't seem right).
Diagnose. Identify which layer failed. Was the instruction unclear (prompt)? Was information missing (context)? Were the wrong tradeoffs made (intent)? Was a risk ignored (judgment)? Is the behavior inconsistent with previous work (coherence)?
Re-enter at the specific layer that needs correction. Not always the prompt. Often the fix is adding a file to context, clarifying a priority, or defining an uncertainty protocol, none of which involve changing the prompt at all.

The resulting skill is mode fluency: recognizing which layer is failing and intervening at the right level. Most AI frustration comes from trying to fix context problems with better prompts, or judgment problems with more context.

§4 component_map

The components in detail

From the Component Reference Map, v1.0, March 2026.

Layer 1

Prompt Engineering

What to do

The craft of designing the structural scaffolding that bridges the gap between a model's theoretical capability and its practical output.

Governs: What to do
Scope: Per-session
Characteristic artifacts: Structured instructions, role definitions, format specs, task decomposition
Failure signature: Incoherent capability: individually plausible, collectively inconsistent

The human feels like they are talking to a brilliant entity that cannot understand what they mean.

Layer 2

Context Engineering

What to know while doing

The architecture of computing the precise slice of state required to focus a model's attention budget on a specific decision.

Governs: What to know
Scope: Per-system
Characteristic artifacts: State files, context documents, retrieval systems, memory stores, compression protocols
Failure signature: Informed incompetence: good instructions, wrong or missing information

The human feels like they are working with someone who follows instructions perfectly but has some sort of troublesome amnesia.

Layer 3

Intent Engineering

What to want while doing

The discipline of making what humans know (goals, values, tradeoffs, decision boundaries) explicit, structured, and machine-actionable.

Governs: What to want
Scope: Per-organization
Characteristic artifacts: Goal hierarchies, values docs, tradeoff specifications, decision boundaries, escalation logic
Failure signature: Misaligned optimization: task done correctly, actual goal undermined

It looks like success until you examine the second-order effects. Metrics improve. Dashboards are green.

Layer 4

Judgment Engineering

What to doubt while doing

The discipline of designing systems that can recognize when their own operating assumptions, including their intent specifications, have become insufficient, and can adapt, escalate, or gracefully degrade rather than optimizing confidently in the wrong direction.

Governs: What to doubt
Scope: Continuous
Characteristic artifacts: Uncertainty protocols, scope boundaries, pre-implementation checkpoints, escalation conditions
Failure signature: Confident wrongness: high metric performance, progressive real-world drift

It hits every metric. It violates no explicit boundary. And it is steadily drifting.

Layer 5

Coherence Engineering

What to become while doing

The discipline of ensuring that a multi-agent ecosystem produces emergent behavior consistent with the overarching purpose of the system, even when individual agents are operating correctly within their own scope.

Governs: What to become
Scope: Per-ecosystem
Characteristic artifacts: Behavioral guidelines, inter-agent protocols, system-level monitoring, architectural principles
Failure signature: Emergent misalignment: each agent correct, collective behavior wrong

Auditing any single agent shows correct behavior. The problem exists only at the system level.

Meta-function

Evaluation Engineering

How to know while doing

Not a layer: a parallel function that observes, measures, and corrects every layer of the stack, continuously and often autonomously. The layers tell you what to configure; evaluation tells you whether it's working.

Governs: How to know
Scope: Cross-layer
Characteristic artifacts: Test suites, assertion libraries, layer-attributed diagnostics, autonomous correction loops
Failure signature: Silent failure: the system can't see what's breaking or where

If you remove evaluation, the layers still function; you just can't see whether they're functioning well.

Meta-function

Harness Engineering

Where and how to do

The infrastructure, tooling, and execution environment in which the stack operates. Not a layer of cognitive concern, but the platform on which all cognitive concerns are implemented.

Governs: Where and how
Scope: Infrastructure
Characteristic artifacts: Execution platform, tool integrations, orchestration definitions, CI/CD pipeline, session management
Failure signature: Capability mismatch: the platform constrains what the cognitive layers could do

The harness is the kitchen; the layers are the recipe. A better kitchen doesn't change the recipe, but it changes what's feasible.

§5 fractal

The fractal property

The same five-layer structure applies at every scale of human-machine collaboration (a property Nate B. Jones identified).

Individual: how you instruct an AI tool, what personal context you provide, what your goals and values are, whether you notice when your approach stops working, and how your various AI tools work together rather than in isolation.
Team: how the team prompts shared AI tools, what organizational data those tools access, what team objectives they serve, whether the team detects when AI outputs are drifting, and how multiple AI-augmented workflows interact.
Organization: the prompt libraries and standards, the data and knowledge architecture, the AI governance framework, the feedback systems that detect misalignment, and the ecosystem management that keeps hundreds of AI deployments coherent.

The individual who masters intent engineering for their personal workflow is developing the exact skills needed to build organizational intent architecture.

Where the regress stops

If judgment monitors intent, what monitors judgment? The regress is terminated by human governance. At some point in every system there must be a human, or a human institution, that says: “This is what we value. This is what we are willing to be held accountable for.” That foundation cannot be automated without circularity.

§6 protocols

The two protocol documents

github · for your agentAI-Human Stack Agent Audit ProtocolFor your agent to run against your system. github · for humansAI-Human Stack Component Reference MapWhat each component looks like when implemented, maturity levels, and an architecture comparison template.

The full documents: The AI-Human Engineering Stack (PDF) and its companion, The Complete Guide to Specifying Work for AI (PDF).

Text on this page is drawn from The AI-Human Engineering Stack and the Component Reference Map (Henrique Jr. Sanchez, March 2026).