Not a composite cognitive units

A team names one callable edge PlanTheWholeWeek. Inside: interpret the request, generate candidates, compose days, critique variety, propose substitutions, and draft shopping notes. Reviewers praise "one clean API." Operators cannot price, substitute, or evaluate any single judgment. When Tuesday fails, the whole call regenerates.

That object bundles several judgments plus orchestration pretending to be atomic, and therefore fails the cognitive unit test regardless of prompt quality. Composite edges feel efficient in demos. They fail in operations, evaluation, and maturity paths. Thoughtware treats them as an anti-pattern with a clear prescription: split judgments, compose in skills or agent strategy, and keep each cognitive unit wrong-answerable on its own terms.

Helpful context: One cognitive unit, one open decision states the rule. Most decisions are several decisions applies the same split instinct upstream. What is a skill packages multi-step known work.

The split prescription

If a workflow contains multiple open judgments with distinct failure modes, the response is to split until each cognitive unit owns one wrong-answerable decision. cognitive units compose in a skill when the procedure shape is known. They wrap with an agent when the path must still be discovered case by case. Enlarging the cognitive unit boundary to match demo convenience is the wrong direction, and prompt length is not proof of atomic judgment.

The negation is firm because composite edges destroy the properties that make Thoughtware workable: locality, substitution, suite binding, cost attribution, and local repair. When any of these properties fails, the system reverts to whole-prompt reasoning about quality, and operators lose the ability to trace a production weakness to a single named decision owner.

Three tests expose a composite. The "and" test checks whether the name contains "and," because GenerateAndCritiquePlan is two cognitive units and orchestration. The failure locality test asks whether operators can say which judgment failed without reading the entire output. The suite binding test asks whether evidence attaches to one wrong-answerable decision, because a suite comment that says "overall quality" likely covers several decisions wearing one badge.

Meal Companion failure mode

PlanTheWholeWeek as one cognitive unit breaks in four predictable ways. Evaluation cannot bind cleanly: variety may be acceptable while Tuesday practicality fails, and a composite suite either scores the whole blob or hides which judgment broke. Substitution becomes rewrite-the-mega-prompt, because teams cannot swap JudgeWeekdayPracticality from the judge library without touching unrelated composition logic. Cost scales with blast radius, since every local fix pays for full replan generation even when one evening needs repair. Authority blurs, because which step required human approval before acting on the world becomes guesswork inside a composite edge.

Decomposed architecture names InterpretWeek, ComposeWeek, CritiquePlan, and judge-library cognitive units such as JudgeWeekdayPracticality. Skills compose the known procedure. The Meal Planning Agent pursues acceptance and chooses transitions after critique.

Split until the edge matches the judgment.

The cognitive unit · Ch. 2

Invoice and enterprise parallels

Invoice extraction tempts the same shortcut: one cognitive unit called ProcessInvoice that reads, interprets, matches, judges, and persists. Field teams want one button. Architecture teams need InterpretHeader, line matchers, materiality judges, and a skill that performs extraction with visible seams. Enterprise systems pay for composite edges in audit findings, rerun cost, and inability to promote one judgment into code when terrain stabilizes.

The same discipline appears in the Decisions track. One arriving question may hide four decisions with three different correct homes. Composite cognitive units are the architecture-layer version of that mistake: selection swallows responsibility inside one callable. Upstream sorting asks where each decision lives. This page asks whether each named edge owns one of those decisions. The answers align.

Why composition is not merger

Skills may contain many cognitive units calls. They must not merge several open judgments into one cognitive unit because composition lives at the skill layer. See a skill may contain cognition. The skill owns procedure. Each cognitive unit still owns one judgment. Confusing composition with merger recreates the composite anti-pattern one level up.

Agents sometimes accumulate composite cognitive edges inside loop bodies because pursuit code grows faster than library discipline. The same split prescription applies: if an agent's inline prompt decides three separable questions, the fix is to extract cognitive units and call them by name. Agent strategy composes cognitive units. It does not replace them with monolithic generation passes that evade suites. The discipline of extraction also prevents loop bodies from becoming untestable: once each judgment has a name, a contract, and a suite pointer, regression tests can target the exact decision that degraded after a model upgrade without replaying an entire pursuit trajectory.

After splitting, suites bind to each cognitive unit and seam tests bind to the skill. A regression where Tuesday fails and variety passes confirms the point: the decomposed system patches locally while the composite system fails the evaluation story loudly enough that nobody ships it again. Shared cognition depends on sealed cognitive units with stable names, and a composite prompt in one product cannot substitute into another team's skill without rewriting hidden judgments nobody named.

Common resistance and the response

Teams resist splitting because demos look slower on day one. Recording the second-week costs changes the picture: full regeneration, unpriceable calls, and missing substitution versus local patch through named cognitive units. The architecture argument is operational, not aesthetic.

Schedule pressure argues for one edge "until later." Later rarely arrives before production traffic. Named cognitive units in the first sprint that touches a judgment, even when templates are thin, create edges that upgrade. Anonymous composite edges ossify. Teams that defer the split often discover that the composite prompt has accumulated implicit dependencies on output formatting, ordering assumptions, and model-specific phrasing that make later extraction expensive enough to justify a full rewrite rather than a surgical split.

Bringing a composite edge to review with three columns, judgment name, failure mode, and suite owner, reveals the blur. Decomposed rows each carry a single wrong-answerable sentence. For Meal Companion, "Tuesday impractical" points to JudgeWeekdayPracticality or AssessMealPracticality, not to an unlabelled stage inside PlanTheWholeWeek. Support tickets read the same way. Finance can ask which judgment drove spend spikes, and named cognitive units answer with engineering facts rather than aggregate numbers.

Leaders sometimes approve composite edges to protect demo dates. A decomposed parallel implementation behind a feature flag, measured against second-week repair cost, usually vindicates split edges without shaming the demo choice. A before-and-after diagram in the architecture wiki anchors the prescription when memory fades between releases.

Composite prompts that bundle several judgments remain tempting because they demo well, and they fail the same way in week three. Splitting compounds early costs less than untangling ownership after launch. Each split should leave a named owner and a suite pointer.

What to do next

The largest callable edge in any AI feature is the starting point. Writing every open decision it makes and splitting when the list exceeds one produces named cognitive units with pointable suites. Known procedures compose in skills, and agent strategy covers discovery. The failure locality test on the decomposed design confirms that local repair paths exist for common weaknesses.

See what is not a cognitive unit and a skill may contain cognition.

Read next: Why pipeline is not enough.