Agent Systems: Structured Output and Dependency Injection

Problem

Free-form language is ambiguous at a workflow boundary. A downstream component should not have to guess whether a sentence is a claim, an error, a citation, or a request for another action.

Intent and structure

task → schema-constrained artifact → parser/validator → next component

The schema can constrain fields, types, enumerated statuses, or a grammar. It does not make the content true; it makes missing or malformed content visible.

Mathematical artifacts

A theorem record can contain objects, hypotheses, conclusion, source, and status. A proof-step record can contain claim, dependencies, justification, verification status, and repair hints. A citation record can contain source, location, extracted statement, and provenance.

Dependency injection

Pass the model, retriever, checker, or tool into a component through an interface. The component then depends on a capability, not on one concrete implementation. This permits offline tests, model replacement, and controlled experiments.

Forces and failure modes

Schemas improve parsing, testing, and auditing, but overly rigid schemas can discard useful uncertainty. Validate both syntax and semantics, and preserve an explicit unknown, incomplete, or needs_review status.

Figure

Figure: a schema boundary makes intermediate artifacts parseable without claiming that they are true.