Summary
The opening article is the map for the series. A supplier invoice arrives looking like a document-processing request, then turns out to combine five different kinds of work: retrieval is an integration problem, interpreting an ambiguous supplier name is a judgement problem, checking thresholds is deterministic, approving an exception is human, and recording the result is an integration problem that must not happen twice.
Opus represents that work as a workflow — a graph of typed task nodes — and one run of it as a case, a distinction that matters because the design is versioned and shared while each case carries its own state, permissions and trace. The shared representation, the Work Knowledge Graph, is introduced carefully: it is a model of what a workflow means rather than a particular database, and it is not a memory store.
The article then walks one case end to end — entry, coordination, waiting, completion — and states the theme the whole series returns to: capturing evidence is not the same as learning from it, and promotion through an explicit gate is what turns a trace into reusable knowledge.
It closes by naming the cross-cutting concerns — identity and authorization, data boundaries, versioning, observability — and by conceding what the alternative composition of best-of-breed tools costs: the integrating team must maintain the links between workflow version, trace, permission and evaluation result. Opus takes on that integration responsibility, and the rest of the series examines one link at a time.