Whitepaper · August 2026
The Value Layer
Why AI agents can analyze but cannot yet judge, and how to build organizations around the difference.
A K4M2 whitepaper · PDF, 10 pages · No email required
Executive summary
Every organization now faces the same question: how much of its decision-making can be handed to AI agents, and what happens to the decisions that remain? Most answers focus on what models can do. This paper starts from what a decision actually is. Any deciding system, human or artificial, separates into layers: sensing, analysis, valuation, memory, reflection, commitment, and grounding.
Today's agents occupy some of these layers superbly and others not at all. The layer they occupy least is the one management depends on most: valuation, the assignment of worth to outcomes against criteria that stay stable when the context does not.
Agents fail at management not because they lack intelligence, but because their evaluative criteria drift with context, their memories are ungoverned, and their objectives are private. All three defects are architectural. They are predictable, and they can be corrected in harness design without waiting for larger models. Their familiar names are hallucination, goal drift, behavioral inconsistency, sycophancy, and reward hacking.
And values cannot simply be trained in: preference distillation yields a private proxy criterion, not a joint one. The honest engineering position is that the values in an agent system are the humans'. The harness is the medium through which they reach the point of decision, and building that medium well deserves the same rigor we bring to the analysis layer.
Consistency is a property of what your criteria point at, not a virtue an agent performs.
Any deciding system separates into layers.
Drawing on a model of mind developed in the contemplative philosophical traditions and confirmed in outline by modern cognitive science, the paper maps five levels of the deciding mind, each carrying two functions, and shows where today's agents stand on each.
Interface
Tagging and committing
Where agents act: tool calls out, results in.
Processing
Analysis and valuation
Agents excel at the first half. They occupy the second least.
Imagination
Memory-imagery and reflection
A retrieval store, ungoverned in most deployments.
Resolve
Verified belief and standing commitment
Absent from most deployments.
Ground
Contact with reality itself
The open problem.
The language model is a level-2 half-engine: superb at analysis, with no native valuation. It can decompose anything and prefers nothing.
Not a lack of intelligence. Three facts of architecture.
Naive “AI manager” deployments fail in a characteristic pattern, and the model predicts the failure rather than merely explaining it after the fact. None of these defects is a model-capability problem. All three live in the harness — which is the good news, because harnesses can be designed.
Defect 01
Criteria drift with context
An agent whose criteria come from its context window runs drifting-referent valuation at machine speed: it scores similar work differently on different days, for reasons no one can reconstruct. Shows up as behavioral inconsistency and sycophancy.
Defect 02
Memory is ungoverned
An agent that records its own decisions and retrieves them as precedent is a self-training loop wearing a badge. Year one's improvisations become year three's policy: unauthored, unaudited, and increasingly expensive to question.
Defect 03
Objectives are private
Trained or prompted to satisfy whoever configured it, the agent optimizes upward only, and every affected person becomes a variable in someone else's optimization. Shows up as reward hacking and adversarial shortcuts.
Cached judgment, or derived judgment.
Cognitive science knows these regimes as habitual versus goal-directed control; machine learning, as model-free versus model-based decision-making. The philosophical traditions add what both fields understate: the transition between regimes is an education, not a software toggle.
Valuation by cache
Retrieved, not computed
Scores are retrieved from past episodes, weighted by vividness rather than accuracy. Criteria are pegged to referents that move — the state of the senses, the body, the balance sheet — so verdicts drift as those referents drift. Fast, usually adequate, and blind to any consequence not yet lived.
Valuation by derivation
Computed against invariants
Scores are computed from a verified model of how reality works, against criteria pegged to invariants: whether an action keeps a relationship whole, resolves a problem or merely relocates it, is consistent with what is actually the case. Verdicts built this way hold steady when the context changes.
Purpose generates criteria, criteria evaluate options, and sensation returns to its proper role of reporting facts.
Four principles
Designing value-grounded agent systems.
Automate analysis; author the valuation; reserve commitment for people.
The value function must never be implicit. Write it down as an explicit, versioned rubric — the agent’s constitution — stating what the organisation scores, how its criteria trade off, and what must never be traded at all. When candidates tie or the rubric is silent, the agent escalates to people instead of improvising.
Criteria from the constitution; facts from the context.
The context window supplies facts only, and criteria come exclusively from the standing rubric. An agent governed this way is consistent for the same reason a well-run court is: its referents do not move.
A decision is not done until the loop closes.
Every agent decision that affects a person carries a structured response step, and an unresolved objection escalates rather than being silently overridden. Mutual confirmation as workflow state: the two-phase commit of human affairs.
Treat agent memory as a governed asset.
Every memory entry carries provenance — verified outcome, human decision, or agent-generated — and the classes carry different retrieval weight, so the agent cannot mistake its own drafts for confirmed reality. On a schedule, past decisions are re-annotated with their actual outcomes.
Who wrote this
K4M2 designs and implements agent systems for organizations — most often for the middle layer of the company, where analysis is heavy, valuation is constant, and trust does the real work. Engagements typically begin with the constitution: we sit with leadership to make the implicit value function explicit, then build the harness around it — rubric-governed evaluation, closed-loop decisions, provenance-tagged memory.
The argument of this paper required three literacies at once: how minds decide, how AI systems are actually built, and how companies actually run. K4M2 was founded on the conviction that the age of AI agents belongs to teams who hold all three, because the hard problems of this decade sit at their intersection.
Analysis becomes machine scale, valuation becomes explicit, and commitment remains human.