Decide, perform, pursue

A team sketches their Meal Companion on a whiteboard. Someone draws three boxes labelled "small LLM calls," "medium workflows," and "big agent." By afternoon the conversation is about token budgets and file counts. Nobody can say who owns whether the weekly plan is acceptable, who runs the critique-and-patch loop, or who decides when to stop. The meeting feels productive because boxes appeared on the board. It produced no clearer ownership.

That sizing meeting is a smell. Thoughtware sorts cognitive work by responsibility, not by scale. Three verbs carry the grammar that the whole skills-and-agents track assumes. Once decide, perform, and pursue are visible on a diagram, questions about model size and line count stop being the main design language.

Helpful context: What is a cognitive unit owns one open judgment. This page names three composition responsibilities that sit above a single cognitive unit. Responsibilities, not sizes explains why line count is the wrong axis for drawing boundaries.

Three verbs, one household

Take the Weekly Meal Companion request: plan dinners for four, Tuesday and Thursday are busy, use the spinach before Wednesday. That sentence hides several kinds of work. Some steps answer point questions under household constraints. Some steps follow a procedure the team already trusts. One module must keep the whole pursuit alive until the plan is accepted or escalated.

Decide belongs to a cognitive unit. AssessMealPracticality judges whether a traybake fits a busy Tuesday evening. The call owns one wrong-answerable question and returns a structured result. It does not compose the week, purchase groceries, or declare the pursuit complete.

Perform belongs to a skill. The Compose Weekly Plan skill assembles a provisional week by calling GenerateCandidates, ComposeWeek, and deterministic validators in a known order. The answers vary with household context, but the shape of the work largely does not. That distinction matters. Performing is not "non-AI." A skill may orchestrate many cognitive units calls and still perform rather than pursue, because its procedure was known before the run began.

Pursue belongs to an agent. The Meal Planning Agent owns the goal of an accepted weekly dinner plan. It iterates critique, patch, fetch, ask, or stop until the minimum sufficient loop is met. Pursuit means goal ownership with working state, loop policy, and designed stopping, not a longer prompt wearing an agent badge.

Deciding is not "smaller" than pursuing. Performing is not a warm-up for agency. Each verb names a different kind of ownership.

A cognitive unit decides. A skill performs. An agent pursues.

The cognitive unit · Ch. 16

Why sizing is the wrong axis

Common misconceptions dissolve once the grammar is visible. Teams say "agents are big skills," but size does not create agency. Goal ownership, review vocabulary, and loop policy do. A compact loop that must revise strategy from evidence remains an agent even when its orchestration code fits on one screen. Teams describe skills as "pipelines without AI," yet skills may orchestrate cognition freely. They package known work, not absence of models. The invoice extraction skill in The cognitive unit calls field-level cognitive units sequentially and still owns one perform responsibility.

The most damaging shortcut is a single mega-prompt that decides the whole week. That collapses several open judgments into one fake edge. Not a composite cognitive units explains why the fix is to split judgments into named cognitive units and compose them in a skill or agent strategy.

When reviews optimize for token spend or file count, they split coherent skills into fragile micro-services or label fixed DAGs as agents because three LLM calls appeared in sequence. Each move hides cost, seams, authority, and evaluation addresses. Responsibility maps replace headcount maps.

Invoice intake uses the same verbs

The Meal Companion is not the only place this grammar appears. Invoice intake in The cognitive unit uses the same three verbs on different terrain.

Field-level cognitive units decide interpretations: what a line item means, whether a discrepancy is material, how variance reads against policy. An extraction skill performs retrieve, validate, and persist as a bounded capability with procedure evaluation. A goal-owning module pursues clearing a queue with review and escalation when cases diverge from the happy path.

The extraction skill may be large in code. It remains perform, not pursue, if the agent or human operator already declared what "done" means for the queue and the skill's job is a known capability inside that pursuit.

Where each verb lives in architecture

Decide lives at the cognitive unit boundary: one judgment, one contract, one suite address. Perform lives at the skill boundary: a named capability, declared inputs and outputs, procedure checks across seams. Pursue lives at the agent boundary: goal, working state, stop conditions, and explicit transitions after each material step.

Deterministic code sits beside all three. Exact checks, permissions, and side effects belong in deterministic code. Code neither decides open questions nor owns pursuit unless a goal has been accidentally smuggled into a script.

1. 01Agent pursuesGoal, loop, delegation
2. 02Skill performsKnown procedure + cognitive unit + code
3. 03cognitive unit decidesOne open judgment
4. 04Code enforcesExact gates and actions

Responsibilities stack. They do not scale by line count.

How the verbs appear in practice

Walk one iteration of the Meal Companion with the three verbs visible. The agent pursues acceptance of the week. Working state holds the interpreted calendar, provisional plan, accepted meals, and open gaps. The agent calls the Compose Weekly Plan skill, which performs candidate generation, composition, and validation. Inside that skill, AssessMealPracticality decides whether each candidate fits its evening. JudgeWeekdayPracticality decides practical fit across the schedule. When critique finds a weakness, the agent chooses patch versus replan versus ask. That choice is pursuit, not perform.

The same grammar transfers to design review. For each box on a diagram, the relevant question is which verb applies. When the answer is decide, a cognitive unit name, contract, and suite address confirm the boundary. When the answer is perform, a skill name and procedure evaluation plan confirm it. When the answer is pursue, a goal sentence, working state sketch, and stop rules confirm it. Boxes that mix verbs signal a boundary problem: a skill that also declares overall success is smuggling pursuit, an agent that re-implements a stable procedure inline is rediscovering perform work every run, and a cognitive unit that composes days and critiques variety is a composite edge waiting for an incident.

The verbs also help product and engineering align on evaluation. Decide work gets decision suites. Perform work gets seam tests and capability metrics. Pursue work gets trajectory review and downstream consequence checks. Mixing those evaluation types produces dashboards that look green while the household still cannot shop with confidence. When a suite reports that every cognitive unit passes yet the overall plan is impractical, the diagnosis is always the same: pursuit evaluation was never designed and trajectory quality went unmeasured while leaf scores gave false comfort.

The three verbs stay useful only when each names a different owner in specs and on-call rotation.

What to do next

The starting point is naming the goal someone must pursue, rather than naming the model stack. From there, listing known procedures that deserve skill packaging prevents the system from rediscovering the same work each run. Remaining open judgments split into cognitive units with contracts and suites, and whiteboard boxes labelled small, medium, and large give way to boxes labelled decide, perform, and pursue.

See what is a skill and what is an agent.

Read next: Responsibilities, not sizes.