Teaching skill composition requires grading the seams, not component recall

Teaching skill composition requires grading the seams, not component recall. A learner can explain every technique in a workflow and still fail when asked to select, order, connect, and recover those techniques under real conditions. Composition is not the sum of isolated scores. It is reliable competence at the boundaries where one skill hands control, state, and assumptions to the next.

The failure taxonomy for composed agent skills is useful because it names the same errors that appear in human performance. Routing failures choose the wrong method for the task. Interface violations pass an output that does not satisfy the next step’s preconditions. Context degradation loses an earlier constraint as the sequence grows. Error cascades consume a plausible but incorrect intermediate result. State corruption makes later reasoning coherent around a false premise. A quiz on definitions sees none of these because it stops before the skills interact.

I would therefore define a composition rubric around transitions. Did the learner recognize which skill applied and why? Did they check its preconditions? Did they transform the intermediate result into the representation the next skill expects? Did they validate before propagating uncertainty? Did they preserve enough state to resume after failure? These are observable behaviors. They turn “can combine skills” from a vague aspiration into evidence that can be inspected.

This extends measuring skill transfer rather than answer throughput. Novel tasks should alter the route, interface, or order while preserving the underlying components. One exercise may supply clean inputs; another should include an ambiguous cue that makes two methods look applicable. One may end successfully; another should inject a failed intermediate step and ask the learner to choose between retry, fallback, compensation, and escalation. Transfer appears when the seams survive variation.

Instruction should expose those seams before assessment does. Worked examples usually compress them into phrases such as “then apply the model” or “pass the result onward.” I want the teacher to stop there: name the contract, show what can be lost, and compare a valid handoff with a plausible invalid one. Progressive complexity should add interface ambiguity and recovery obligations, not merely larger inputs. Complexity that only increases volume trains endurance rather than composition.

Feedback also has to localize failure. Marking the final answer wrong encourages the learner to repeat the whole chain. A seam-level diagnosis can say that the selected components were sound but the second transition dropped a unit, assumed idempotency, or accepted an unverified result. That distinction matters because retries without diagnosis create error accumulation. Calibration and escalation belong in the rubric: recognizing that a transition is unsafe is often better evidence than confidently completing it.

There is one precise concession: early instruction should still assess component recall and fluency. Learners cannot compose primitives they do not possess, and excessive boundary complexity can overload working memory before foundations stabilize. The limit is temporal—component tests are scaffolding, not proof of end-to-end competence.

The final assessment should make the joins visible in the scoring even when the finished artifact looks correct. A lucky output can conceal a broken route; a minor final error can conceal strong recovery reasoning. Grading the seams rewards the capabilities that survive unfamiliar tools, changed order, and partial failure. That is where compositional skill actually lives—and where both people and agents most often lose it.