Pipelines are not agents

Most teams build a pipeline first: extract, classify, assess, draft, in a straight line. That shape is legitimate. It is often the right first ship. Calling it an agent obscures what is missing: stop conditions, revision, authority, and terrain-sensitive depth. Marketing renames the chain. Architecture reviews inherit a false sense of completeness, and incident response discovers that nobody designed what happens when the fixed path encounters a case it cannot handle.

Helpful context: Why pipeline is not enough explains what DAGs omit. What is an agent adds goal pursuit and loops. What is a skill packages known capabilities.

Invoice intake: useful and still not an agent

Four hundred thousand invoices arrive. Most are routine. Exceptions follow extract, match, classify, priority, draft response. There is no loop. Nothing revises strategy mid-case. Nothing owns queue clearance as a goal with escalation when shape diverges. The work is measurable, valuable, and architecturally a pipeline.

When that work becomes a named capability with a contract, inspectable dependencies, and procedure-level evaluation, the architectural name is Skill. Pipeline names an implementation shape. Skill names a reusable capability the organisation can search, substitute, and evaluate as a procedure. Teams resist this label because pipeline sounds smaller than agent. Smaller is sometimes correct. A pipeline that composes named cognitive units, carries evaluation suites, and keeps closed work in code can be production-grade. It still fails the agent test when case shape varies and a fixed path handles unfamiliar shapes badly. The distinction is not about ambition or complexity: a pipeline processing four hundred thousand routine invoices with named cognitive units at each interpretation stage delivers real production value without needing pursuit vocabulary, because the work genuinely follows a fixed path for every case.

What an agent adds

An agent owns a goal whose path may not be known in advance. It selects capabilities, observes results, updates working state, and chooses an explicit transition: accept, patch, fetch, bridge, replan, ask, escalate, or stop.

Property	Pipeline or Skill	Agent
Path	Substantially known at design time	Chosen per case
Stopping	End of the procedure	Designed stop conditions
Revision	Bounded branches inside the procedure	Local repair and replan
Authority	Skills make grants explicit	Explicit grants and escalation
Cost	Predictable per stage	Bounded by policy and terrain

The Meal Companion contrast makes the table concrete. Recipe fetch, plan text, display is a pipeline. Critique, patch, targeted question, and escalation until an accepted plan emerges is a loop. Same domain. Different obligations. Walk one lap of agent pursuit. CritiquePlan names a weekday weakness. The agent selects local patch via RecommendMealSubstitution, full replan when the week representation fails, or ASK when household preference is missing. A pipeline cannot choose among those transitions because nothing owns pursuit after critique returns. The pipeline version would output the first plan it produced regardless of quality, and the household would bear the full cost of discovering its impracticality at the grocery store rather than during a review loop that costs only model tokens.

Useful does not mean agent

Pipelines belong inside Skills as perform steps. Agents own goals, not every stage in a chain. Confusing the two leads to two failures. The first is false completeness: stakeholders believe autonomy exists because the diagram has many boxes, and incidents reveal no stop design, no escalation, no working state. The second is missing Skill packaging: teams leave valuable fixed paths as tribal scripts because "agent" felt like the only respectable label, and procedure evaluation never arrives.

This is a pipeline and not an agent, since there is no loop and nothing decides when to stop.

The cognitive unit Ch. 15

When unfamiliar shapes become common, composition becomes a loop. That is a promotion decision based on evidence, not on marketing vocabulary. A fixed acting chain with no loop remains a pipeline or acting Skill depending on whether it carries a capability contract. Calling it an agent because it touches production systems is one of the five things called agent.

Two shapes in one product

Even within the Meal Companion, not everything is an agent lap. Candidate generation inside a settled week may follow a known Skill path: call GenerateCandidates, invoke ComposeWeek, run deterministic validators, return a provisional artifact. That perform work is valuable precisely because it is repeatable. Agent work begins when evidence can change what should happen next: after critique, after a failed guard, after a missing preference, after a novel medical restriction returns leadership to the household. See decide, perform, pursue.

Real products combine shapes. An agent may call Skills that internally look like pipelines. A pipeline may call cognitive units with open judgments inside fixed stages. The label applies to the outer pursuit module, not to every nested graph. The Meal Planning Agent orchestrates loops. Inside it, Compose Weekly Plan may run a substantially fixed subgraph. That nesting is healthy. Confusion arrives when the outer module lacks goal ownership because the inner graph was mistaken for the whole architecture. The nesting also clarifies evaluation responsibility: the inner pipeline evaluates as a skill with seam tests, while the outer agent evaluates as pursuit with trajectory quality and stop discipline, and neither evaluation program substitutes for the other.

Questions before shipping the label

Four questions resolve the debate in design review. Does anything decide when to stop based on outcome, not stage count? Can the path revise without restarting the whole chain from scratch? Are authority and escalation explicit before side effects run? Is there working state beyond chat history? If any answer is no, the system likely has a pipeline or Skill. Naming it accurately attracts the right evaluation and ownership. Accurate naming is a favour to future you. Incident response begins with "what architecture did we claim?" If the claim was agent and the build was pipeline, every hour of debugging starts with vocabulary repair. Product teams that label pipelines honestly find their evaluation programs become more tractable, because pipeline evaluation tests stage outputs and throughput rather than attempting trajectory scoring on a system that has no trajectory to score.

Evaluation differs by shape

Pipeline evaluation focuses on stage outputs and throughput. Skill evaluation adds procedure seams and recovery. Agent evaluation adds trajectory quality and stop discipline. Mislabeling shapes assigns the wrong evaluation program. Teams then argue about fluency when the defect is missing stop design. Name the shape before naming the metric. A pipeline KPI on an agent loop hides pursuit failures. An agent trajectory review on a stable Skill wastes effort. Getting the label right on day one saves the evaluation team from building the wrong test harness and prevents months of dashboards that measure the wrong properties of the system.

Cognitive orchestration coordinates agents, Skills, cognitive units, guards, and memory toward a goal. Pipelines may appear inside orchestration as perform steps. They do not replace orchestration when pursuit is required. Minimum sufficient loop applies to agents, not to pipelines. A pipeline ends when its last stage completes. An agent stops when sufficiency, authority, and terrain say so.

Marketing labels change slowly, but architecture reviews can still ask whether the loop owns a goal or merely transforms inputs.

What to do next

Redraw the flagship flow and highlight return paths and review decisions. If none exist, the slide title "agent" describes intent rather than architecture, and honest naming opens the path to correct evaluation. Package stable paths as Skills with procedure evaluation. Promote to an agent only when case shape and revision requirements justify pursuit. Keep a running list of promoted shapes, since promotion should follow evidence from production traces rather than demo enthusiasm.

See when a trajectory becomes a Skill, minimum sufficient loop, and acting is not agency.

Read next: Cognitive orchestration.