Opus.com · Three posts
Three posts written for the Opus.com blog over a week in September 2026. They run from a problem — a workflow whose every step is valid and which has still lost any route to an acceptable finished design — to the structure Opus keeps its process knowledge in, and then to the people and agents who maintain that structure. Each has its own page here, and each of those links to the post on Opus.com.
Read in this order
These are listed oldest first, which is not the usual way round for a blog and is deliberate. The three were published within a week and they build on each other: the first argues that validity at every step is not the same as a workflow that can still be finished, the second describes the Work Knowledge Graph that holds the tasks and connections such a workflow is assembled from, and the third is about who maintains that graph and what a proposed addition has to survive. Meeting Graph Gardening before the graph it gardens would be the wrong way in. The dates are on each entry, so the order is visible rather than assumed.
—
A workflow can be correct at every step and still have lost every route to an acceptable finished design.
Every task is valid and every connection permitted, and the design has still painted itself into a corner. Whether an addition is safe turns out to depend on the obligations the workflow already carries and the options it has left, which is not a property of the addition.
The Work Knowledge Graph keeps reusable tasks, the connections between them and the evidence behind both, so a workflow starts from knowledge already established.
Two valid steps do not automatically form a valid process, so the evidence has to support the handoffs as well as the tasks. The graph also preserves the differences — industry, geography, language, process detail — that decide what can safely be shared.
Gardeners find the gaps and prepare the additions, and every candidate has to pass grounding, structure, entity resolution and approval before it lands.
An Arabic-language retail-banking onboarding example carries the post. Autonomous gardeners gather sources and propose tasks and connections, and what happens to a candidate after that is the substance.
One thread runs through all three. A workflow is not made trustworthy by the correctness of its parts: the first post makes that argument about the steps of a single design, the second makes it about the evidence behind a shared graph of tasks, and the third makes it about the additions that graph accepts, where a new connection between workflows needs a human and a new task may not. Each post also names what it does not deliver — a certificate that does not promise completion, a graph that does not promise execution, checks that can be wrong — which is the more useful half of each claim. The Inside Opus series covers the running system these posts sit underneath, and the papers give the formal treatment.
—