Summary
Graph Gardening is how the Work Knowledge Graph is maintained and extended: identifying the gaps in it and preparing additions supported by evidence. The post follows a single worked example throughout — onboarding for Arabic-language retail banking — which keeps the machinery attached to a process someone actually has to run.
Autonomous gardeners gather sources and propose tasks and connections, and the proposal is the easy half. Each candidate then passes through factual grounding, structural validation, entity resolution and approval. Where policy permits, eligible additions may be approved automatically; new connections between workflows require human review. That asymmetry is the post’s sharpest point, and it follows from the previous two: the connections are where a set of individually defensible steps can still fail to compose.
Approved changes enter immutable snapshots that retain the evidence and the version history, so an addition can be traced back to what justified it. Concurrent updates are governed by conflict checks and revalidation, and private enterprise graphs keep records of their own and require explicit client authorisation before anything contributes back to the master graph.
The practical limits are named too, and they are the sort that are usually left out: evidence checks can be wrong, spending checks do not guarantee against overspend, and where the graph is stored does not determine where model inference runs. What the process is claimed to produce is knowledge whose additions can be inspected, traced and corrected — not knowledge that is certainly right.