Architectural Overview

The Structural Problem

Engineering organizations adopting AI face a structural mismatch.

• Generation is cheap — AI produces output at a scale and speed that has no precedent.
• Absorption is expensive — humans review, integrate, govern, and learn from output at human speed.
• Cost compounds with usage — agentic systems running in loops produce surprise bills that pause adoption.
• Decisions accumulate without traceability — outputs ship without audit trails.
• Learnings stay trapped — what team A discovered, team B reinvents.

Without architectural intervention, the typical outcome is sophisticated per-team agentic systems whose gains do not compound, whose costs are invisible, and whose decisions cannot be audited. KCC is the operating model designed to address each of these failure modes structurally.

4.1 The Architectural Diagram

The arrows are not runtime invocation paths. They are inheritance and influence relationships: the kernel defines the contract capabilities and cells inherit; the catalog is composed of L2/L3 capabilities available to all cells; meta-agents act on every invocation; and the Inspector Pipeline observes traces from all cells and proposes promotions back into the catalog.

4.2 Runtime Flow

A typical agent invocation flows as follows.

• A user or upstream agent makes a request to a specific agent in a cell.
• The cell's agent declaration is loaded, including its full agent contract.
• The Token Guard inspects the request against the cost envelope and blast-radius declarations. Refuses if exceeded.
• The Butler scores the request for value, risk, complexity, confidence, reviewer attention, and trifecta status. Decides HITL or HOTL.
• If HITL: the request is queued for human review before the agent runs.
• If HOTL or after approval: the agent runs against its declared tools and within its cost envelope.
• The Token Guard monitors during execution and cancels if hard ceilings are breached.
• The agent produces output with confidence and decision trace.
• Telemetry is emitted and recorded in the Inspector Pipeline's observation stream.
• The output is returned to the caller.

A single invocation, end to end:

This flow is not centralized through the kernel at runtime. The kernel's role is specification, not orchestration. Each cell may implement the flow with different infrastructure as long as the contract is honored.

4.3 The Compounding Mechanism

The architecture's central claim is that learning compounds across teams. This happens through three mechanisms.

• The Shared Kernel — all cells implement the same contract; when the kernel is improved, all cells benefit.
• The Capability Catalog — a useful agent that generalizes can be promoted and adopted by other cells when they are ready.
• The Inspector Pipeline — patterns observed in one cell's traces become proposals that affect all cells.

These three mechanisms together transform per-team innovation into organizational intelligence over time. Without them, the architecture is just per-team agentic adoption with extra ceremony.

4.4 What Sits Where

The kernel changes least often. Capabilities change moderately. Cells change continuously. The three change cadences permit different concerns to evolve at appropriate speeds.

The Nine Surfaces

Every agent declares nine surfaces. The first seven are operational — they govern what the agent does. The last two are cognitive — they make agent decisions observable, replayable, and patternable.

The cognitive layer is the most distinctive technical contribution of KCC. Surfaces 8 and 9 together form the cognitive substrate: the infrastructure through which an organization remembers what its agents decided and why.

The Three Layers