Two different boundaries

Two cuts confuse even careful teams, and the confusion shows up in the same diagram meeting.

The cognitive unit boundary asks which single open judgment this callable edge owns. The skill boundary asks which known capability this reusable procedure performs. Conflating them produces either composite cognitive edges that hide several judgments or skills that smuggle several responsibilities behind one name.

Both cuts are necessary. They answer different questions. Getting both visible on the same architecture drawing is the "aha" this page targets.

Helpful context: One cognitive unit, one open decision defines the first cut. What is a skill defines the second. Responsibilities, not sizes warns against line-count thinking.

Side by side

The cognitive unit cut answers: what single judgment is this? AssessMealPracticality decides whether a candidate meal fits a busy Tuesday evening. Callers invoke that judgment without reading the full meal-planning procedure.

The skill cut answers: what known capability is this? Compose Weekly Plan orchestrates candidates, composition, and validation into one perform responsibility. Callers invoke the capability without re-specifying every cognitive unit call inside it.

These cuts are orthogonal. The same cognitive unit may appear in two skills without merging judgments. Practicality assessment can serve both Compose Weekly Plan and Repair One Evening because both capabilities need the same point judgment. That reuse is only possible when the cognitive unit boundary was drawn at the judgment level rather than at the skill level, because a judgment packaged correctly travels across capabilities without duplication or drift.

Conversely, one skill contains two cognitive units without becoming a composite cognitive units, because the skill owns procedure, not a larger merged judgment. Composition is not merger.

cognitive unit Acognitive unit BSkill boundaryCallable capability

cognitive units split judgments. Skills package procedures that call them.

Two diagrams, one product

Picture two diagrams for the Meal Companion.

Diagram one: reuse across skills. AssessMealPracticality sits in a library box with a suite pointer. An arrow leads to Compose Weekly Plan. Another arrow leads to Repair One Evening. Same judgment, two capabilities, no prompt duplication.

Diagram two: composition inside one skill. Compose Weekly Plan contains GenerateCandidates, ComposeWeek, and deterministic validation in known order. Several cognitive units calls, one perform responsibility, procedure tests on seams.

Teams that draw only diagram two sometimes merge cognitive units to "simplify" the skill innards. Teams that draw only diagram one sometimes forget procedure ownership and let agents rediscover ordering every run. Both lenses are necessary.

A third boundary appears when cognition travels beyond one product: libraries of cognition package sealed cognitive units and skills for substitution across teams. The cognitive unit cut still applies inside packages, because each exported edge still owns one judgment. The library cut applies when assets cross team boundaries with version pins and substitution rules. Confusing library packaging with cognitive unit merger is a common enterprise mistake. A library skill is still one capability. A library cognitive units is still one judgment.

The cognitive unit boundary separates judgments. The skill boundary separates responsibilities.

The cognitive unit · Ch. 16

The aha test

Two questions on every box in the diagram expose the distinction. Who owns this judgment? If the box decides an open question, it belongs as a cognitive unit with a contract and suite. Who owns this capability? If the box performs known work by composing cognitive units and code, it belongs as a skill with procedure evaluation.

When both answers differ, both boundaries need to be visible. When both answers feel true for the same box, a composite edge or a skill that smuggles pursuit is the likely cause.

Agents add a third responsibility cut: goal pursuit. Agents compose skills and cognitive units through working state and loop policy, but they do not erase the first two boundaries by inlining every judgment and procedure into one loop body. See where the cognitive unit ends and the agent begins.

Common conflation mistakes follow predictable patterns. Renaming a skill after one cognitive unit inside it, for example "MaterialitySkill" that does nothing beyond judging materiality, usually means a cognitive unit has been mispackaged. Declaring a procedure as one cognitive unit because callers want one invoke mistakes convenience for architecture. Drawing cognitive unit boxes inside agent loops with no skill layer for stable procedures turns agents into mega-scripts where perform work never accrues procedure evidence.

Versioning, substitution, and portability

The cognitive unit boundary enables substitution: swap AssessMealPracticality v3 without rewriting Compose Weekly Plan procedure order. The skill boundary enables capability versioning: Compose Weekly Plan v2 adds a validation seam without merging header interpretation into line matching.

When both boundaries are visible, release notes speak precisely. "Updated practicality judge" and "updated compose capability" are different communications with different regression suites.

When cognition travels in libraries of cognition, the cognitive unit cut travels inside packages and the skill cut travels as named capabilities. Products differ in pursuit policies while sharing perform and decide assets. That portability fails when boundaries were never drawn clearly in the first place.

Invoice parallels reinforce the point. Header interpretation is a cognitive unit. Line matching is another. Extraction as a capability is a skill. Clearing the invoice queue with escalation is pursuit. Swapping a field interpreter in a library updates one judgment owner without rewriting extraction procedure order. Changing extraction procedure order updates one skill owner without merging header interpretation into line matching. That independence is the payoff of drawing both boundaries clearly: each owner can change at its own pace without cascading rewrites into adjacent concerns.

Both cuts on every diagram

Agent diagrams still need both cuts visible. Agents compose skills and cognitive units, and they do not erase the two boundaries by collapsing everything into loop prose. If the agent box hides unnamed cognition, composite edges inside pursuit code are the likely explanation. If the agent box re-implements a skill inline, perform work is being rediscovered every run.

New teammates often conflate "one invoke" with "one judgment." The two-question test on their first diagram review builds the habit. Within a month the reach for cognitive unit names when a box decides, and skill names when a box performs, becomes natural. The habit also improves code review: when a pull request introduces a new callable, reviewers instinctively ask whether it owns a judgment or a procedure, and that question prevents compounds from entering the codebase under convenience names.

Pull request reviews carry the same discipline. Asking whether any box both decides and performs, whether any skill name hides two capabilities, and whether any cognitive unit name hides an "and" catches merges that feel convenient. Drawing Repair One Evening beside Compose Weekly Plan when teaching reuse shows the same practicality cognitive unit serving both capabilities. That single diagram often convinces teams faster than another paragraph about orthogonality.

Diagrams that show both cuts reduce arguments about whether a box is too big or merely mislabeled. Whiteboard arguments shrink when both cuts appear on the same diagram.

What to do next

The exercise is drawing both boundaries on the same diagram: cognitive units for judgments, skills for capabilities. Reusing cognitive units across skills where the same judgment appears and composing multiple cognitive units inside skills without merging judgments makes both cuts concrete. Stable agent trajectories promote into named skills over time.

See why pipeline is not enough and where the cognitive unit ends and the agent begins.

Read next: Acting is not agency.