#Graphs #AgentWorkflows #Orchestration #NewRuntime
The most useful point in the Graphs batch is not that the word “graph” replaced the word “loop.” The loop is still there: act -> inspect -> correct remains a normal local mechanism. But once there are multiple outcomes, shared state, approvals, recovery, and several workers, the loop stops describing the system.
A graph is needed when the execution path itself becomes a product surface. One item can become post, watch, blocked_missing_primary_source, or deep_research. One PR can move through cohorts, layers, stale snapshots, and review threads. One desktop agent can reach send, write, or command only through an approval gate.
So I would read “graphs are the new loops” this way: loops stay inside nodes, while the graph decides who is allowed to move to the next state. It is not a pretty diagram. It is a governance layer.
In New Runtime, this is already operational reality: source resolution, draft, visual review, site, R2, OpenClaw receipt, and owner approval are not one cycle. They are a graph with receipts at external write boundaries. The next step is to make that graph explicit in data, not only in the runbook.